1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This contains code to emit Builtin calls as LLVM code.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "CGCXXABI.h"
14 #include "CGObjCRuntime.h"
15 #include "CGOpenCLRuntime.h"
16 #include "CGRecordLayout.h"
17 #include "CodeGenFunction.h"
18 #include "CodeGenModule.h"
19 #include "ConstantEmitter.h"
20 #include "PatternInit.h"
21 #include "TargetInfo.h"
22 #include "clang/AST/ASTContext.h"
23 #include "clang/AST/Attr.h"
24 #include "clang/AST/Decl.h"
25 #include "clang/AST/OSLog.h"
26 #include "clang/Basic/TargetBuiltins.h"
27 #include "clang/Basic/TargetInfo.h"
28 #include "clang/CodeGen/CGFunctionInfo.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/StringExtras.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/Support/ConvertUTF.h"
36 #include "llvm/Support/ScopedPrinter.h"
37 #include "llvm/Support/TargetParser.h"
38 #include <sstream>
39 
40 using namespace clang;
41 using namespace CodeGen;
42 using namespace llvm;
43 
44 static
45 int64_t clamp(int64_t Value, int64_t Low, int64_t High) {
46   return std::min(High, std::max(Low, Value));
47 }
48 
49 static void initializeAlloca(CodeGenFunction &CGF, AllocaInst *AI, Value *Size, unsigned AlignmentInBytes) {
50   ConstantInt *Byte;
51   switch (CGF.getLangOpts().getTrivialAutoVarInit()) {
52   case LangOptions::TrivialAutoVarInitKind::Uninitialized:
53     // Nothing to initialize.
54     return;
55   case LangOptions::TrivialAutoVarInitKind::Zero:
56     Byte = CGF.Builder.getInt8(0x00);
57     break;
58   case LangOptions::TrivialAutoVarInitKind::Pattern: {
59     llvm::Type *Int8 = llvm::IntegerType::getInt8Ty(CGF.CGM.getLLVMContext());
60     Byte = llvm::dyn_cast<llvm::ConstantInt>(
61         initializationPatternFor(CGF.CGM, Int8));
62     break;
63   }
64   }
65   CGF.Builder.CreateMemSet(AI, Byte, Size, AlignmentInBytes);
66 }
67 
68 /// getBuiltinLibFunction - Given a builtin id for a function like
69 /// "__builtin_fabsf", return a Function* for "fabsf".
70 llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
71                                                      unsigned BuiltinID) {
72   assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
73 
74   // Get the name, skip over the __builtin_ prefix (if necessary).
75   StringRef Name;
76   GlobalDecl D(FD);
77 
78   // If the builtin has been declared explicitly with an assembler label,
79   // use the mangled name. This differs from the plain label on platforms
80   // that prefix labels.
81   if (FD->hasAttr<AsmLabelAttr>())
82     Name = getMangledName(D);
83   else
84     Name = Context.BuiltinInfo.getName(BuiltinID) + 10;
85 
86   llvm::FunctionType *Ty =
87     cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
88 
89   return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
90 }
91 
92 /// Emit the conversions required to turn the given value into an
93 /// integer of the given size.
94 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
95                         QualType T, llvm::IntegerType *IntType) {
96   V = CGF.EmitToMemory(V, T);
97 
98   if (V->getType()->isPointerTy())
99     return CGF.Builder.CreatePtrToInt(V, IntType);
100 
101   assert(V->getType() == IntType);
102   return V;
103 }
104 
105 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
106                           QualType T, llvm::Type *ResultType) {
107   V = CGF.EmitFromMemory(V, T);
108 
109   if (ResultType->isPointerTy())
110     return CGF.Builder.CreateIntToPtr(V, ResultType);
111 
112   assert(V->getType() == ResultType);
113   return V;
114 }
115 
116 /// Utility to insert an atomic instruction based on Intrinsic::ID
117 /// and the expression node.
118 static Value *MakeBinaryAtomicValue(
119     CodeGenFunction &CGF, llvm::AtomicRMWInst::BinOp Kind, const CallExpr *E,
120     AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
121   QualType T = E->getType();
122   assert(E->getArg(0)->getType()->isPointerType());
123   assert(CGF.getContext().hasSameUnqualifiedType(T,
124                                   E->getArg(0)->getType()->getPointeeType()));
125   assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
126 
127   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
128   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
129 
130   llvm::IntegerType *IntType =
131     llvm::IntegerType::get(CGF.getLLVMContext(),
132                            CGF.getContext().getTypeSize(T));
133   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
134 
135   llvm::Value *Args[2];
136   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
137   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
138   llvm::Type *ValueType = Args[1]->getType();
139   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
140 
141   llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
142       Kind, Args[0], Args[1], Ordering);
143   return EmitFromInt(CGF, Result, T, ValueType);
144 }
145 
146 static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) {
147   Value *Val = CGF.EmitScalarExpr(E->getArg(0));
148   Value *Address = CGF.EmitScalarExpr(E->getArg(1));
149 
150   // Convert the type of the pointer to a pointer to the stored type.
151   Val = CGF.EmitToMemory(Val, E->getArg(0)->getType());
152   Value *BC = CGF.Builder.CreateBitCast(
153       Address, llvm::PointerType::getUnqual(Val->getType()), "cast");
154   LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType());
155   LV.setNontemporal(true);
156   CGF.EmitStoreOfScalar(Val, LV, false);
157   return nullptr;
158 }
159 
160 static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) {
161   Value *Address = CGF.EmitScalarExpr(E->getArg(0));
162 
163   LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType());
164   LV.setNontemporal(true);
165   return CGF.EmitLoadOfScalar(LV, E->getExprLoc());
166 }
167 
168 static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
169                                llvm::AtomicRMWInst::BinOp Kind,
170                                const CallExpr *E) {
171   return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E));
172 }
173 
174 /// Utility to insert an atomic instruction based Intrinsic::ID and
175 /// the expression node, where the return value is the result of the
176 /// operation.
177 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
178                                    llvm::AtomicRMWInst::BinOp Kind,
179                                    const CallExpr *E,
180                                    Instruction::BinaryOps Op,
181                                    bool Invert = false) {
182   QualType T = E->getType();
183   assert(E->getArg(0)->getType()->isPointerType());
184   assert(CGF.getContext().hasSameUnqualifiedType(T,
185                                   E->getArg(0)->getType()->getPointeeType()));
186   assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
187 
188   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
189   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
190 
191   llvm::IntegerType *IntType =
192     llvm::IntegerType::get(CGF.getLLVMContext(),
193                            CGF.getContext().getTypeSize(T));
194   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
195 
196   llvm::Value *Args[2];
197   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
198   llvm::Type *ValueType = Args[1]->getType();
199   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
200   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
201 
202   llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
203       Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent);
204   Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
205   if (Invert)
206     Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result,
207                                      llvm::ConstantInt::get(IntType, -1));
208   Result = EmitFromInt(CGF, Result, T, ValueType);
209   return RValue::get(Result);
210 }
211 
212 /// Utility to insert an atomic cmpxchg instruction.
213 ///
214 /// @param CGF The current codegen function.
215 /// @param E   Builtin call expression to convert to cmpxchg.
216 ///            arg0 - address to operate on
217 ///            arg1 - value to compare with
218 ///            arg2 - new value
219 /// @param ReturnBool Specifies whether to return success flag of
220 ///                   cmpxchg result or the old value.
221 ///
222 /// @returns result of cmpxchg, according to ReturnBool
223 ///
224 /// Note: In order to lower Microsoft's _InterlockedCompareExchange* intrinsics
225 /// invoke the function EmitAtomicCmpXchgForMSIntrin.
226 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E,
227                                      bool ReturnBool) {
228   QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType();
229   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
230   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
231 
232   llvm::IntegerType *IntType = llvm::IntegerType::get(
233       CGF.getLLVMContext(), CGF.getContext().getTypeSize(T));
234   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
235 
236   Value *Args[3];
237   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
238   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
239   llvm::Type *ValueType = Args[1]->getType();
240   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
241   Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType);
242 
243   Value *Pair = CGF.Builder.CreateAtomicCmpXchg(
244       Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent,
245       llvm::AtomicOrdering::SequentiallyConsistent);
246   if (ReturnBool)
247     // Extract boolean success flag and zext it to int.
248     return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1),
249                                   CGF.ConvertType(E->getType()));
250   else
251     // Extract old value and emit it using the same type as compare value.
252     return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T,
253                        ValueType);
254 }
255 
256 /// This function should be invoked to emit atomic cmpxchg for Microsoft's
257 /// _InterlockedCompareExchange* intrinsics which have the following signature:
258 /// T _InterlockedCompareExchange(T volatile *Destination,
259 ///                               T Exchange,
260 ///                               T Comparand);
261 ///
262 /// Whereas the llvm 'cmpxchg' instruction has the following syntax:
263 /// cmpxchg *Destination, Comparand, Exchange.
264 /// So we need to swap Comparand and Exchange when invoking
265 /// CreateAtomicCmpXchg. That is the reason we could not use the above utility
266 /// function MakeAtomicCmpXchgValue since it expects the arguments to be
267 /// already swapped.
268 
269 static
270 Value *EmitAtomicCmpXchgForMSIntrin(CodeGenFunction &CGF, const CallExpr *E,
271     AtomicOrdering SuccessOrdering = AtomicOrdering::SequentiallyConsistent) {
272   assert(E->getArg(0)->getType()->isPointerType());
273   assert(CGF.getContext().hasSameUnqualifiedType(
274       E->getType(), E->getArg(0)->getType()->getPointeeType()));
275   assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),
276                                                  E->getArg(1)->getType()));
277   assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),
278                                                  E->getArg(2)->getType()));
279 
280   auto *Destination = CGF.EmitScalarExpr(E->getArg(0));
281   auto *Comparand = CGF.EmitScalarExpr(E->getArg(2));
282   auto *Exchange = CGF.EmitScalarExpr(E->getArg(1));
283 
284   // For Release ordering, the failure ordering should be Monotonic.
285   auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release ?
286                          AtomicOrdering::Monotonic :
287                          SuccessOrdering;
288 
289   auto *Result = CGF.Builder.CreateAtomicCmpXchg(
290                    Destination, Comparand, Exchange,
291                    SuccessOrdering, FailureOrdering);
292   Result->setVolatile(true);
293   return CGF.Builder.CreateExtractValue(Result, 0);
294 }
295 
296 static Value *EmitAtomicIncrementValue(CodeGenFunction &CGF, const CallExpr *E,
297     AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
298   assert(E->getArg(0)->getType()->isPointerType());
299 
300   auto *IntTy = CGF.ConvertType(E->getType());
301   auto *Result = CGF.Builder.CreateAtomicRMW(
302                    AtomicRMWInst::Add,
303                    CGF.EmitScalarExpr(E->getArg(0)),
304                    ConstantInt::get(IntTy, 1),
305                    Ordering);
306   return CGF.Builder.CreateAdd(Result, ConstantInt::get(IntTy, 1));
307 }
308 
309 static Value *EmitAtomicDecrementValue(CodeGenFunction &CGF, const CallExpr *E,
310     AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
311   assert(E->getArg(0)->getType()->isPointerType());
312 
313   auto *IntTy = CGF.ConvertType(E->getType());
314   auto *Result = CGF.Builder.CreateAtomicRMW(
315                    AtomicRMWInst::Sub,
316                    CGF.EmitScalarExpr(E->getArg(0)),
317                    ConstantInt::get(IntTy, 1),
318                    Ordering);
319   return CGF.Builder.CreateSub(Result, ConstantInt::get(IntTy, 1));
320 }
321 
322 // Build a plain volatile load.
323 static Value *EmitISOVolatileLoad(CodeGenFunction &CGF, const CallExpr *E) {
324   Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
325   QualType ElTy = E->getArg(0)->getType()->getPointeeType();
326   CharUnits LoadSize = CGF.getContext().getTypeSizeInChars(ElTy);
327   llvm::Type *ITy =
328       llvm::IntegerType::get(CGF.getLLVMContext(), LoadSize.getQuantity() * 8);
329   Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
330   llvm::LoadInst *Load = CGF.Builder.CreateAlignedLoad(Ptr, LoadSize);
331   Load->setVolatile(true);
332   return Load;
333 }
334 
335 // Build a plain volatile store.
336 static Value *EmitISOVolatileStore(CodeGenFunction &CGF, const CallExpr *E) {
337   Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
338   Value *Value = CGF.EmitScalarExpr(E->getArg(1));
339   QualType ElTy = E->getArg(0)->getType()->getPointeeType();
340   CharUnits StoreSize = CGF.getContext().getTypeSizeInChars(ElTy);
341   llvm::Type *ITy =
342       llvm::IntegerType::get(CGF.getLLVMContext(), StoreSize.getQuantity() * 8);
343   Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
344   llvm::StoreInst *Store =
345       CGF.Builder.CreateAlignedStore(Value, Ptr, StoreSize);
346   Store->setVolatile(true);
347   return Store;
348 }
349 
350 // Emit a simple mangled intrinsic that has 1 argument and a return type
351 // matching the argument type.
352 static Value *emitUnaryBuiltin(CodeGenFunction &CGF,
353                                const CallExpr *E,
354                                unsigned IntrinsicID) {
355   llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
356 
357   Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
358   return CGF.Builder.CreateCall(F, Src0);
359 }
360 
361 // Emit an intrinsic that has 2 operands of the same type as its result.
362 static Value *emitBinaryBuiltin(CodeGenFunction &CGF,
363                                 const CallExpr *E,
364                                 unsigned IntrinsicID) {
365   llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
366   llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
367 
368   Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
369   return CGF.Builder.CreateCall(F, { Src0, Src1 });
370 }
371 
372 // Emit an intrinsic that has 3 operands of the same type as its result.
373 static Value *emitTernaryBuiltin(CodeGenFunction &CGF,
374                                  const CallExpr *E,
375                                  unsigned IntrinsicID) {
376   llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
377   llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
378   llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
379 
380   Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
381   return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 });
382 }
383 
384 // Emit an intrinsic that has 1 float or double operand, and 1 integer.
385 static Value *emitFPIntBuiltin(CodeGenFunction &CGF,
386                                const CallExpr *E,
387                                unsigned IntrinsicID) {
388   llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
389   llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
390 
391   Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
392   return CGF.Builder.CreateCall(F, {Src0, Src1});
393 }
394 
395 // Emit an intrinsic that has overloaded integer result and fp operand.
396 static Value *emitFPToIntRoundBuiltin(CodeGenFunction &CGF,
397                                       const CallExpr *E,
398                                       unsigned IntrinsicID) {
399    llvm::Type *ResultType = CGF.ConvertType(E->getType());
400    llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
401 
402    Function *F = CGF.CGM.getIntrinsic(IntrinsicID,
403                                       {ResultType, Src0->getType()});
404    return CGF.Builder.CreateCall(F, Src0);
405 }
406 
407 /// EmitFAbs - Emit a call to @llvm.fabs().
408 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) {
409   Function *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType());
410   llvm::CallInst *Call = CGF.Builder.CreateCall(F, V);
411   Call->setDoesNotAccessMemory();
412   return Call;
413 }
414 
415 /// Emit the computation of the sign bit for a floating point value. Returns
416 /// the i1 sign bit value.
417 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) {
418   LLVMContext &C = CGF.CGM.getLLVMContext();
419 
420   llvm::Type *Ty = V->getType();
421   int Width = Ty->getPrimitiveSizeInBits();
422   llvm::Type *IntTy = llvm::IntegerType::get(C, Width);
423   V = CGF.Builder.CreateBitCast(V, IntTy);
424   if (Ty->isPPC_FP128Ty()) {
425     // We want the sign bit of the higher-order double. The bitcast we just
426     // did works as if the double-double was stored to memory and then
427     // read as an i128. The "store" will put the higher-order double in the
428     // lower address in both little- and big-Endian modes, but the "load"
429     // will treat those bits as a different part of the i128: the low bits in
430     // little-Endian, the high bits in big-Endian. Therefore, on big-Endian
431     // we need to shift the high bits down to the low before truncating.
432     Width >>= 1;
433     if (CGF.getTarget().isBigEndian()) {
434       Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width);
435       V = CGF.Builder.CreateLShr(V, ShiftCst);
436     }
437     // We are truncating value in order to extract the higher-order
438     // double, which we will be using to extract the sign from.
439     IntTy = llvm::IntegerType::get(C, Width);
440     V = CGF.Builder.CreateTrunc(V, IntTy);
441   }
442   Value *Zero = llvm::Constant::getNullValue(IntTy);
443   return CGF.Builder.CreateICmpSLT(V, Zero);
444 }
445 
446 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *FD,
447                               const CallExpr *E, llvm::Constant *calleeValue) {
448   CGCallee callee = CGCallee::forDirect(calleeValue, GlobalDecl(FD));
449   return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot());
450 }
451 
452 /// Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
453 /// depending on IntrinsicID.
454 ///
455 /// \arg CGF The current codegen function.
456 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
457 /// \arg X The first argument to the llvm.*.with.overflow.*.
458 /// \arg Y The second argument to the llvm.*.with.overflow.*.
459 /// \arg Carry The carry returned by the llvm.*.with.overflow.*.
460 /// \returns The result (i.e. sum/product) returned by the intrinsic.
461 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF,
462                                           const llvm::Intrinsic::ID IntrinsicID,
463                                           llvm::Value *X, llvm::Value *Y,
464                                           llvm::Value *&Carry) {
465   // Make sure we have integers of the same width.
466   assert(X->getType() == Y->getType() &&
467          "Arguments must be the same type. (Did you forget to make sure both "
468          "arguments have the same integer width?)");
469 
470   Function *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
471   llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y});
472   Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
473   return CGF.Builder.CreateExtractValue(Tmp, 0);
474 }
475 
476 static Value *emitRangedBuiltin(CodeGenFunction &CGF,
477                                 unsigned IntrinsicID,
478                                 int low, int high) {
479     llvm::MDBuilder MDHelper(CGF.getLLVMContext());
480     llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high));
481     Function *F = CGF.CGM.getIntrinsic(IntrinsicID, {});
482     llvm::Instruction *Call = CGF.Builder.CreateCall(F);
483     Call->setMetadata(llvm::LLVMContext::MD_range, RNode);
484     return Call;
485 }
486 
487 namespace {
488   struct WidthAndSignedness {
489     unsigned Width;
490     bool Signed;
491   };
492 }
493 
494 static WidthAndSignedness
495 getIntegerWidthAndSignedness(const clang::ASTContext &context,
496                              const clang::QualType Type) {
497   assert(Type->isIntegerType() && "Given type is not an integer.");
498   unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width;
499   bool Signed = Type->isSignedIntegerType();
500   return {Width, Signed};
501 }
502 
503 // Given one or more integer types, this function produces an integer type that
504 // encompasses them: any value in one of the given types could be expressed in
505 // the encompassing type.
506 static struct WidthAndSignedness
507 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) {
508   assert(Types.size() > 0 && "Empty list of types.");
509 
510   // If any of the given types is signed, we must return a signed type.
511   bool Signed = false;
512   for (const auto &Type : Types) {
513     Signed |= Type.Signed;
514   }
515 
516   // The encompassing type must have a width greater than or equal to the width
517   // of the specified types.  Additionally, if the encompassing type is signed,
518   // its width must be strictly greater than the width of any unsigned types
519   // given.
520   unsigned Width = 0;
521   for (const auto &Type : Types) {
522     unsigned MinWidth = Type.Width + (Signed && !Type.Signed);
523     if (Width < MinWidth) {
524       Width = MinWidth;
525     }
526   }
527 
528   return {Width, Signed};
529 }
530 
531 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) {
532   llvm::Type *DestType = Int8PtrTy;
533   if (ArgValue->getType() != DestType)
534     ArgValue =
535         Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data());
536 
537   Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend;
538   return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue);
539 }
540 
541 /// Checks if using the result of __builtin_object_size(p, @p From) in place of
542 /// __builtin_object_size(p, @p To) is correct
543 static bool areBOSTypesCompatible(int From, int To) {
544   // Note: Our __builtin_object_size implementation currently treats Type=0 and
545   // Type=2 identically. Encoding this implementation detail here may make
546   // improving __builtin_object_size difficult in the future, so it's omitted.
547   return From == To || (From == 0 && To == 1) || (From == 3 && To == 2);
548 }
549 
550 static llvm::Value *
551 getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) {
552   return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true);
553 }
554 
555 llvm::Value *
556 CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type,
557                                                  llvm::IntegerType *ResType,
558                                                  llvm::Value *EmittedE,
559                                                  bool IsDynamic) {
560   uint64_t ObjectSize;
561   if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type))
562     return emitBuiltinObjectSize(E, Type, ResType, EmittedE, IsDynamic);
563   return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true);
564 }
565 
566 /// Returns a Value corresponding to the size of the given expression.
567 /// This Value may be either of the following:
568 ///   - A llvm::Argument (if E is a param with the pass_object_size attribute on
569 ///     it)
570 ///   - A call to the @llvm.objectsize intrinsic
571 ///
572 /// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null
573 /// and we wouldn't otherwise try to reference a pass_object_size parameter,
574 /// we'll call @llvm.objectsize on EmittedE, rather than emitting E.
575 llvm::Value *
576 CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type,
577                                        llvm::IntegerType *ResType,
578                                        llvm::Value *EmittedE, bool IsDynamic) {
579   // We need to reference an argument if the pointer is a parameter with the
580   // pass_object_size attribute.
581   if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) {
582     auto *Param = dyn_cast<ParmVarDecl>(D->getDecl());
583     auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>();
584     if (Param != nullptr && PS != nullptr &&
585         areBOSTypesCompatible(PS->getType(), Type)) {
586       auto Iter = SizeArguments.find(Param);
587       assert(Iter != SizeArguments.end());
588 
589       const ImplicitParamDecl *D = Iter->second;
590       auto DIter = LocalDeclMap.find(D);
591       assert(DIter != LocalDeclMap.end());
592 
593       return EmitLoadOfScalar(DIter->second, /*Volatile=*/false,
594                               getContext().getSizeType(), E->getBeginLoc());
595     }
596   }
597 
598   // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't
599   // evaluate E for side-effects. In either case, we shouldn't lower to
600   // @llvm.objectsize.
601   if (Type == 3 || (!EmittedE && E->HasSideEffects(getContext())))
602     return getDefaultBuiltinObjectSizeResult(Type, ResType);
603 
604   Value *Ptr = EmittedE ? EmittedE : EmitScalarExpr(E);
605   assert(Ptr->getType()->isPointerTy() &&
606          "Non-pointer passed to __builtin_object_size?");
607 
608   Function *F =
609       CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()});
610 
611   // LLVM only supports 0 and 2, make sure that we pass along that as a boolean.
612   Value *Min = Builder.getInt1((Type & 2) != 0);
613   // For GCC compatibility, __builtin_object_size treat NULL as unknown size.
614   Value *NullIsUnknown = Builder.getTrue();
615   Value *Dynamic = Builder.getInt1(IsDynamic);
616   return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown, Dynamic});
617 }
618 
619 namespace {
620 /// A struct to generically describe a bit test intrinsic.
621 struct BitTest {
622   enum ActionKind : uint8_t { TestOnly, Complement, Reset, Set };
623   enum InterlockingKind : uint8_t {
624     Unlocked,
625     Sequential,
626     Acquire,
627     Release,
628     NoFence
629   };
630 
631   ActionKind Action;
632   InterlockingKind Interlocking;
633   bool Is64Bit;
634 
635   static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
636 };
637 } // namespace
638 
639 BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
640   switch (BuiltinID) {
641     // Main portable variants.
642   case Builtin::BI_bittest:
643     return {TestOnly, Unlocked, false};
644   case Builtin::BI_bittestandcomplement:
645     return {Complement, Unlocked, false};
646   case Builtin::BI_bittestandreset:
647     return {Reset, Unlocked, false};
648   case Builtin::BI_bittestandset:
649     return {Set, Unlocked, false};
650   case Builtin::BI_interlockedbittestandreset:
651     return {Reset, Sequential, false};
652   case Builtin::BI_interlockedbittestandset:
653     return {Set, Sequential, false};
654 
655     // X86-specific 64-bit variants.
656   case Builtin::BI_bittest64:
657     return {TestOnly, Unlocked, true};
658   case Builtin::BI_bittestandcomplement64:
659     return {Complement, Unlocked, true};
660   case Builtin::BI_bittestandreset64:
661     return {Reset, Unlocked, true};
662   case Builtin::BI_bittestandset64:
663     return {Set, Unlocked, true};
664   case Builtin::BI_interlockedbittestandreset64:
665     return {Reset, Sequential, true};
666   case Builtin::BI_interlockedbittestandset64:
667     return {Set, Sequential, true};
668 
669     // ARM/AArch64-specific ordering variants.
670   case Builtin::BI_interlockedbittestandset_acq:
671     return {Set, Acquire, false};
672   case Builtin::BI_interlockedbittestandset_rel:
673     return {Set, Release, false};
674   case Builtin::BI_interlockedbittestandset_nf:
675     return {Set, NoFence, false};
676   case Builtin::BI_interlockedbittestandreset_acq:
677     return {Reset, Acquire, false};
678   case Builtin::BI_interlockedbittestandreset_rel:
679     return {Reset, Release, false};
680   case Builtin::BI_interlockedbittestandreset_nf:
681     return {Reset, NoFence, false};
682   }
683   llvm_unreachable("expected only bittest intrinsics");
684 }
685 
686 static char bitActionToX86BTCode(BitTest::ActionKind A) {
687   switch (A) {
688   case BitTest::TestOnly:   return '\0';
689   case BitTest::Complement: return 'c';
690   case BitTest::Reset:      return 'r';
691   case BitTest::Set:        return 's';
692   }
693   llvm_unreachable("invalid action");
694 }
695 
696 static llvm::Value *EmitX86BitTestIntrinsic(CodeGenFunction &CGF,
697                                             BitTest BT,
698                                             const CallExpr *E, Value *BitBase,
699                                             Value *BitPos) {
700   char Action = bitActionToX86BTCode(BT.Action);
701   char SizeSuffix = BT.Is64Bit ? 'q' : 'l';
702 
703   // Build the assembly.
704   SmallString<64> Asm;
705   raw_svector_ostream AsmOS(Asm);
706   if (BT.Interlocking != BitTest::Unlocked)
707     AsmOS << "lock ";
708   AsmOS << "bt";
709   if (Action)
710     AsmOS << Action;
711   AsmOS << SizeSuffix << " $2, ($1)\n\tsetc ${0:b}";
712 
713   // Build the constraints. FIXME: We should support immediates when possible.
714   std::string Constraints = "=r,r,r,~{cc},~{flags},~{fpsr}";
715   llvm::IntegerType *IntType = llvm::IntegerType::get(
716       CGF.getLLVMContext(),
717       CGF.getContext().getTypeSize(E->getArg(1)->getType()));
718   llvm::Type *IntPtrType = IntType->getPointerTo();
719   llvm::FunctionType *FTy =
720       llvm::FunctionType::get(CGF.Int8Ty, {IntPtrType, IntType}, false);
721 
722   llvm::InlineAsm *IA =
723       llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
724   return CGF.Builder.CreateCall(IA, {BitBase, BitPos});
725 }
726 
727 static llvm::AtomicOrdering
728 getBitTestAtomicOrdering(BitTest::InterlockingKind I) {
729   switch (I) {
730   case BitTest::Unlocked:   return llvm::AtomicOrdering::NotAtomic;
731   case BitTest::Sequential: return llvm::AtomicOrdering::SequentiallyConsistent;
732   case BitTest::Acquire:    return llvm::AtomicOrdering::Acquire;
733   case BitTest::Release:    return llvm::AtomicOrdering::Release;
734   case BitTest::NoFence:    return llvm::AtomicOrdering::Monotonic;
735   }
736   llvm_unreachable("invalid interlocking");
737 }
738 
739 /// Emit a _bittest* intrinsic. These intrinsics take a pointer to an array of
740 /// bits and a bit position and read and optionally modify the bit at that
741 /// position. The position index can be arbitrarily large, i.e. it can be larger
742 /// than 31 or 63, so we need an indexed load in the general case.
743 static llvm::Value *EmitBitTestIntrinsic(CodeGenFunction &CGF,
744                                          unsigned BuiltinID,
745                                          const CallExpr *E) {
746   Value *BitBase = CGF.EmitScalarExpr(E->getArg(0));
747   Value *BitPos = CGF.EmitScalarExpr(E->getArg(1));
748 
749   BitTest BT = BitTest::decodeBitTestBuiltin(BuiltinID);
750 
751   // X86 has special BT, BTC, BTR, and BTS instructions that handle the array
752   // indexing operation internally. Use them if possible.
753   llvm::Triple::ArchType Arch = CGF.getTarget().getTriple().getArch();
754   if (Arch == llvm::Triple::x86 || Arch == llvm::Triple::x86_64)
755     return EmitX86BitTestIntrinsic(CGF, BT, E, BitBase, BitPos);
756 
757   // Otherwise, use generic code to load one byte and test the bit. Use all but
758   // the bottom three bits as the array index, and the bottom three bits to form
759   // a mask.
760   // Bit = BitBaseI8[BitPos >> 3] & (1 << (BitPos & 0x7)) != 0;
761   Value *ByteIndex = CGF.Builder.CreateAShr(
762       BitPos, llvm::ConstantInt::get(BitPos->getType(), 3), "bittest.byteidx");
763   Value *BitBaseI8 = CGF.Builder.CreatePointerCast(BitBase, CGF.Int8PtrTy);
764   Address ByteAddr(CGF.Builder.CreateInBoundsGEP(CGF.Int8Ty, BitBaseI8,
765                                                  ByteIndex, "bittest.byteaddr"),
766                    CharUnits::One());
767   Value *PosLow =
768       CGF.Builder.CreateAnd(CGF.Builder.CreateTrunc(BitPos, CGF.Int8Ty),
769                             llvm::ConstantInt::get(CGF.Int8Ty, 0x7));
770 
771   // The updating instructions will need a mask.
772   Value *Mask = nullptr;
773   if (BT.Action != BitTest::TestOnly) {
774     Mask = CGF.Builder.CreateShl(llvm::ConstantInt::get(CGF.Int8Ty, 1), PosLow,
775                                  "bittest.mask");
776   }
777 
778   // Check the action and ordering of the interlocked intrinsics.
779   llvm::AtomicOrdering Ordering = getBitTestAtomicOrdering(BT.Interlocking);
780 
781   Value *OldByte = nullptr;
782   if (Ordering != llvm::AtomicOrdering::NotAtomic) {
783     // Emit a combined atomicrmw load/store operation for the interlocked
784     // intrinsics.
785     llvm::AtomicRMWInst::BinOp RMWOp = llvm::AtomicRMWInst::Or;
786     if (BT.Action == BitTest::Reset) {
787       Mask = CGF.Builder.CreateNot(Mask);
788       RMWOp = llvm::AtomicRMWInst::And;
789     }
790     OldByte = CGF.Builder.CreateAtomicRMW(RMWOp, ByteAddr.getPointer(), Mask,
791                                           Ordering);
792   } else {
793     // Emit a plain load for the non-interlocked intrinsics.
794     OldByte = CGF.Builder.CreateLoad(ByteAddr, "bittest.byte");
795     Value *NewByte = nullptr;
796     switch (BT.Action) {
797     case BitTest::TestOnly:
798       // Don't store anything.
799       break;
800     case BitTest::Complement:
801       NewByte = CGF.Builder.CreateXor(OldByte, Mask);
802       break;
803     case BitTest::Reset:
804       NewByte = CGF.Builder.CreateAnd(OldByte, CGF.Builder.CreateNot(Mask));
805       break;
806     case BitTest::Set:
807       NewByte = CGF.Builder.CreateOr(OldByte, Mask);
808       break;
809     }
810     if (NewByte)
811       CGF.Builder.CreateStore(NewByte, ByteAddr);
812   }
813 
814   // However we loaded the old byte, either by plain load or atomicrmw, shift
815   // the bit into the low position and mask it to 0 or 1.
816   Value *ShiftedByte = CGF.Builder.CreateLShr(OldByte, PosLow, "bittest.shr");
817   return CGF.Builder.CreateAnd(
818       ShiftedByte, llvm::ConstantInt::get(CGF.Int8Ty, 1), "bittest.res");
819 }
820 
821 namespace {
822 enum class MSVCSetJmpKind {
823   _setjmpex,
824   _setjmp3,
825   _setjmp
826 };
827 }
828 
829 /// MSVC handles setjmp a bit differently on different platforms. On every
830 /// architecture except 32-bit x86, the frame address is passed. On x86, extra
831 /// parameters can be passed as variadic arguments, but we always pass none.
832 static RValue EmitMSVCRTSetJmp(CodeGenFunction &CGF, MSVCSetJmpKind SJKind,
833                                const CallExpr *E) {
834   llvm::Value *Arg1 = nullptr;
835   llvm::Type *Arg1Ty = nullptr;
836   StringRef Name;
837   bool IsVarArg = false;
838   if (SJKind == MSVCSetJmpKind::_setjmp3) {
839     Name = "_setjmp3";
840     Arg1Ty = CGF.Int32Ty;
841     Arg1 = llvm::ConstantInt::get(CGF.IntTy, 0);
842     IsVarArg = true;
843   } else {
844     Name = SJKind == MSVCSetJmpKind::_setjmp ? "_setjmp" : "_setjmpex";
845     Arg1Ty = CGF.Int8PtrTy;
846     if (CGF.getTarget().getTriple().getArch() == llvm::Triple::aarch64) {
847       Arg1 = CGF.Builder.CreateCall(
848           CGF.CGM.getIntrinsic(Intrinsic::sponentry, CGF.AllocaInt8PtrTy));
849     } else
850       Arg1 = CGF.Builder.CreateCall(
851           CGF.CGM.getIntrinsic(Intrinsic::frameaddress, CGF.AllocaInt8PtrTy),
852           llvm::ConstantInt::get(CGF.Int32Ty, 0));
853   }
854 
855   // Mark the call site and declaration with ReturnsTwice.
856   llvm::Type *ArgTypes[2] = {CGF.Int8PtrTy, Arg1Ty};
857   llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get(
858       CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex,
859       llvm::Attribute::ReturnsTwice);
860   llvm::FunctionCallee SetJmpFn = CGF.CGM.CreateRuntimeFunction(
861       llvm::FunctionType::get(CGF.IntTy, ArgTypes, IsVarArg), Name,
862       ReturnsTwiceAttr, /*Local=*/true);
863 
864   llvm::Value *Buf = CGF.Builder.CreateBitOrPointerCast(
865       CGF.EmitScalarExpr(E->getArg(0)), CGF.Int8PtrTy);
866   llvm::Value *Args[] = {Buf, Arg1};
867   llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(SetJmpFn, Args);
868   CB->setAttributes(ReturnsTwiceAttr);
869   return RValue::get(CB);
870 }
871 
872 // Many of MSVC builtins are on x64, ARM and AArch64; to avoid repeating code,
873 // we handle them here.
874 enum class CodeGenFunction::MSVCIntrin {
875   _BitScanForward,
876   _BitScanReverse,
877   _InterlockedAnd,
878   _InterlockedDecrement,
879   _InterlockedExchange,
880   _InterlockedExchangeAdd,
881   _InterlockedExchangeSub,
882   _InterlockedIncrement,
883   _InterlockedOr,
884   _InterlockedXor,
885   _InterlockedExchangeAdd_acq,
886   _InterlockedExchangeAdd_rel,
887   _InterlockedExchangeAdd_nf,
888   _InterlockedExchange_acq,
889   _InterlockedExchange_rel,
890   _InterlockedExchange_nf,
891   _InterlockedCompareExchange_acq,
892   _InterlockedCompareExchange_rel,
893   _InterlockedCompareExchange_nf,
894   _InterlockedOr_acq,
895   _InterlockedOr_rel,
896   _InterlockedOr_nf,
897   _InterlockedXor_acq,
898   _InterlockedXor_rel,
899   _InterlockedXor_nf,
900   _InterlockedAnd_acq,
901   _InterlockedAnd_rel,
902   _InterlockedAnd_nf,
903   _InterlockedIncrement_acq,
904   _InterlockedIncrement_rel,
905   _InterlockedIncrement_nf,
906   _InterlockedDecrement_acq,
907   _InterlockedDecrement_rel,
908   _InterlockedDecrement_nf,
909   __fastfail,
910 };
911 
912 Value *CodeGenFunction::EmitMSVCBuiltinExpr(MSVCIntrin BuiltinID,
913                                             const CallExpr *E) {
914   switch (BuiltinID) {
915   case MSVCIntrin::_BitScanForward:
916   case MSVCIntrin::_BitScanReverse: {
917     Value *ArgValue = EmitScalarExpr(E->getArg(1));
918 
919     llvm::Type *ArgType = ArgValue->getType();
920     llvm::Type *IndexType =
921       EmitScalarExpr(E->getArg(0))->getType()->getPointerElementType();
922     llvm::Type *ResultType = ConvertType(E->getType());
923 
924     Value *ArgZero = llvm::Constant::getNullValue(ArgType);
925     Value *ResZero = llvm::Constant::getNullValue(ResultType);
926     Value *ResOne = llvm::ConstantInt::get(ResultType, 1);
927 
928     BasicBlock *Begin = Builder.GetInsertBlock();
929     BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn);
930     Builder.SetInsertPoint(End);
931     PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result");
932 
933     Builder.SetInsertPoint(Begin);
934     Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero);
935     BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn);
936     Builder.CreateCondBr(IsZero, End, NotZero);
937     Result->addIncoming(ResZero, Begin);
938 
939     Builder.SetInsertPoint(NotZero);
940     Address IndexAddress = EmitPointerWithAlignment(E->getArg(0));
941 
942     if (BuiltinID == MSVCIntrin::_BitScanForward) {
943       Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
944       Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
945       ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
946       Builder.CreateStore(ZeroCount, IndexAddress, false);
947     } else {
948       unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth();
949       Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1);
950 
951       Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
952       Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
953       ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
954       Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount);
955       Builder.CreateStore(Index, IndexAddress, false);
956     }
957     Builder.CreateBr(End);
958     Result->addIncoming(ResOne, NotZero);
959 
960     Builder.SetInsertPoint(End);
961     return Result;
962   }
963   case MSVCIntrin::_InterlockedAnd:
964     return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E);
965   case MSVCIntrin::_InterlockedExchange:
966     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E);
967   case MSVCIntrin::_InterlockedExchangeAdd:
968     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E);
969   case MSVCIntrin::_InterlockedExchangeSub:
970     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E);
971   case MSVCIntrin::_InterlockedOr:
972     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E);
973   case MSVCIntrin::_InterlockedXor:
974     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E);
975   case MSVCIntrin::_InterlockedExchangeAdd_acq:
976     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
977                                  AtomicOrdering::Acquire);
978   case MSVCIntrin::_InterlockedExchangeAdd_rel:
979     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
980                                  AtomicOrdering::Release);
981   case MSVCIntrin::_InterlockedExchangeAdd_nf:
982     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
983                                  AtomicOrdering::Monotonic);
984   case MSVCIntrin::_InterlockedExchange_acq:
985     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
986                                  AtomicOrdering::Acquire);
987   case MSVCIntrin::_InterlockedExchange_rel:
988     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
989                                  AtomicOrdering::Release);
990   case MSVCIntrin::_InterlockedExchange_nf:
991     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
992                                  AtomicOrdering::Monotonic);
993   case MSVCIntrin::_InterlockedCompareExchange_acq:
994     return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Acquire);
995   case MSVCIntrin::_InterlockedCompareExchange_rel:
996     return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Release);
997   case MSVCIntrin::_InterlockedCompareExchange_nf:
998     return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Monotonic);
999   case MSVCIntrin::_InterlockedOr_acq:
1000     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1001                                  AtomicOrdering::Acquire);
1002   case MSVCIntrin::_InterlockedOr_rel:
1003     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1004                                  AtomicOrdering::Release);
1005   case MSVCIntrin::_InterlockedOr_nf:
1006     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1007                                  AtomicOrdering::Monotonic);
1008   case MSVCIntrin::_InterlockedXor_acq:
1009     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1010                                  AtomicOrdering::Acquire);
1011   case MSVCIntrin::_InterlockedXor_rel:
1012     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1013                                  AtomicOrdering::Release);
1014   case MSVCIntrin::_InterlockedXor_nf:
1015     return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1016                                  AtomicOrdering::Monotonic);
1017   case MSVCIntrin::_InterlockedAnd_acq:
1018     return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1019                                  AtomicOrdering::Acquire);
1020   case MSVCIntrin::_InterlockedAnd_rel:
1021     return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1022                                  AtomicOrdering::Release);
1023   case MSVCIntrin::_InterlockedAnd_nf:
1024     return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1025                                  AtomicOrdering::Monotonic);
1026   case MSVCIntrin::_InterlockedIncrement_acq:
1027     return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Acquire);
1028   case MSVCIntrin::_InterlockedIncrement_rel:
1029     return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Release);
1030   case MSVCIntrin::_InterlockedIncrement_nf:
1031     return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Monotonic);
1032   case MSVCIntrin::_InterlockedDecrement_acq:
1033     return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Acquire);
1034   case MSVCIntrin::_InterlockedDecrement_rel:
1035     return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Release);
1036   case MSVCIntrin::_InterlockedDecrement_nf:
1037     return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Monotonic);
1038 
1039   case MSVCIntrin::_InterlockedDecrement:
1040     return EmitAtomicDecrementValue(*this, E);
1041   case MSVCIntrin::_InterlockedIncrement:
1042     return EmitAtomicIncrementValue(*this, E);
1043 
1044   case MSVCIntrin::__fastfail: {
1045     // Request immediate process termination from the kernel. The instruction
1046     // sequences to do this are documented on MSDN:
1047     // https://msdn.microsoft.com/en-us/library/dn774154.aspx
1048     llvm::Triple::ArchType ISA = getTarget().getTriple().getArch();
1049     StringRef Asm, Constraints;
1050     switch (ISA) {
1051     default:
1052       ErrorUnsupported(E, "__fastfail call for this architecture");
1053       break;
1054     case llvm::Triple::x86:
1055     case llvm::Triple::x86_64:
1056       Asm = "int $$0x29";
1057       Constraints = "{cx}";
1058       break;
1059     case llvm::Triple::thumb:
1060       Asm = "udf #251";
1061       Constraints = "{r0}";
1062       break;
1063     case llvm::Triple::aarch64:
1064       Asm = "brk #0xF003";
1065       Constraints = "{w0}";
1066     }
1067     llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false);
1068     llvm::InlineAsm *IA =
1069         llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
1070     llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
1071         getLLVMContext(), llvm::AttributeList::FunctionIndex,
1072         llvm::Attribute::NoReturn);
1073     llvm::CallInst *CI = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0)));
1074     CI->setAttributes(NoReturnAttr);
1075     return CI;
1076   }
1077   }
1078   llvm_unreachable("Incorrect MSVC intrinsic!");
1079 }
1080 
1081 namespace {
1082 // ARC cleanup for __builtin_os_log_format
1083 struct CallObjCArcUse final : EHScopeStack::Cleanup {
1084   CallObjCArcUse(llvm::Value *object) : object(object) {}
1085   llvm::Value *object;
1086 
1087   void Emit(CodeGenFunction &CGF, Flags flags) override {
1088     CGF.EmitARCIntrinsicUse(object);
1089   }
1090 };
1091 }
1092 
1093 Value *CodeGenFunction::EmitCheckedArgForBuiltin(const Expr *E,
1094                                                  BuiltinCheckKind Kind) {
1095   assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero)
1096           && "Unsupported builtin check kind");
1097 
1098   Value *ArgValue = EmitScalarExpr(E);
1099   if (!SanOpts.has(SanitizerKind::Builtin) || !getTarget().isCLZForZeroUndef())
1100     return ArgValue;
1101 
1102   SanitizerScope SanScope(this);
1103   Value *Cond = Builder.CreateICmpNE(
1104       ArgValue, llvm::Constant::getNullValue(ArgValue->getType()));
1105   EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin),
1106             SanitizerHandler::InvalidBuiltin,
1107             {EmitCheckSourceLocation(E->getExprLoc()),
1108              llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)},
1109             None);
1110   return ArgValue;
1111 }
1112 
1113 /// Get the argument type for arguments to os_log_helper.
1114 static CanQualType getOSLogArgType(ASTContext &C, int Size) {
1115   QualType UnsignedTy = C.getIntTypeForBitwidth(Size * 8, /*Signed=*/false);
1116   return C.getCanonicalType(UnsignedTy);
1117 }
1118 
1119 llvm::Function *CodeGenFunction::generateBuiltinOSLogHelperFunction(
1120     const analyze_os_log::OSLogBufferLayout &Layout,
1121     CharUnits BufferAlignment) {
1122   ASTContext &Ctx = getContext();
1123 
1124   llvm::SmallString<64> Name;
1125   {
1126     raw_svector_ostream OS(Name);
1127     OS << "__os_log_helper";
1128     OS << "_" << BufferAlignment.getQuantity();
1129     OS << "_" << int(Layout.getSummaryByte());
1130     OS << "_" << int(Layout.getNumArgsByte());
1131     for (const auto &Item : Layout.Items)
1132       OS << "_" << int(Item.getSizeByte()) << "_"
1133          << int(Item.getDescriptorByte());
1134   }
1135 
1136   if (llvm::Function *F = CGM.getModule().getFunction(Name))
1137     return F;
1138 
1139   llvm::SmallVector<QualType, 4> ArgTys;
1140   FunctionArgList Args;
1141   Args.push_back(ImplicitParamDecl::Create(
1142       Ctx, nullptr, SourceLocation(), &Ctx.Idents.get("buffer"), Ctx.VoidPtrTy,
1143       ImplicitParamDecl::Other));
1144   ArgTys.emplace_back(Ctx.VoidPtrTy);
1145 
1146   for (unsigned int I = 0, E = Layout.Items.size(); I < E; ++I) {
1147     char Size = Layout.Items[I].getSizeByte();
1148     if (!Size)
1149       continue;
1150 
1151     QualType ArgTy = getOSLogArgType(Ctx, Size);
1152     Args.push_back(ImplicitParamDecl::Create(
1153         Ctx, nullptr, SourceLocation(),
1154         &Ctx.Idents.get(std::string("arg") + llvm::to_string(I)), ArgTy,
1155         ImplicitParamDecl::Other));
1156     ArgTys.emplace_back(ArgTy);
1157   }
1158 
1159   QualType ReturnTy = Ctx.VoidTy;
1160   QualType FuncionTy = Ctx.getFunctionType(ReturnTy, ArgTys, {});
1161 
1162   // The helper function has linkonce_odr linkage to enable the linker to merge
1163   // identical functions. To ensure the merging always happens, 'noinline' is
1164   // attached to the function when compiling with -Oz.
1165   const CGFunctionInfo &FI =
1166       CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, Args);
1167   llvm::FunctionType *FuncTy = CGM.getTypes().GetFunctionType(FI);
1168   llvm::Function *Fn = llvm::Function::Create(
1169       FuncTy, llvm::GlobalValue::LinkOnceODRLinkage, Name, &CGM.getModule());
1170   Fn->setVisibility(llvm::GlobalValue::HiddenVisibility);
1171   CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn);
1172   CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn);
1173   Fn->setDoesNotThrow();
1174 
1175   // Attach 'noinline' at -Oz.
1176   if (CGM.getCodeGenOpts().OptimizeSize == 2)
1177     Fn->addFnAttr(llvm::Attribute::NoInline);
1178 
1179   auto NL = ApplyDebugLocation::CreateEmpty(*this);
1180   IdentifierInfo *II = &Ctx.Idents.get(Name);
1181   FunctionDecl *FD = FunctionDecl::Create(
1182       Ctx, Ctx.getTranslationUnitDecl(), SourceLocation(), SourceLocation(), II,
1183       FuncionTy, nullptr, SC_PrivateExtern, false, false);
1184 
1185   StartFunction(FD, ReturnTy, Fn, FI, Args);
1186 
1187   // Create a scope with an artificial location for the body of this function.
1188   auto AL = ApplyDebugLocation::CreateArtificial(*this);
1189 
1190   CharUnits Offset;
1191   Address BufAddr(Builder.CreateLoad(GetAddrOfLocalVar(Args[0]), "buf"),
1192                   BufferAlignment);
1193   Builder.CreateStore(Builder.getInt8(Layout.getSummaryByte()),
1194                       Builder.CreateConstByteGEP(BufAddr, Offset++, "summary"));
1195   Builder.CreateStore(Builder.getInt8(Layout.getNumArgsByte()),
1196                       Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs"));
1197 
1198   unsigned I = 1;
1199   for (const auto &Item : Layout.Items) {
1200     Builder.CreateStore(
1201         Builder.getInt8(Item.getDescriptorByte()),
1202         Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor"));
1203     Builder.CreateStore(
1204         Builder.getInt8(Item.getSizeByte()),
1205         Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize"));
1206 
1207     CharUnits Size = Item.size();
1208     if (!Size.getQuantity())
1209       continue;
1210 
1211     Address Arg = GetAddrOfLocalVar(Args[I]);
1212     Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset, "argData");
1213     Addr = Builder.CreateBitCast(Addr, Arg.getPointer()->getType(),
1214                                  "argDataCast");
1215     Builder.CreateStore(Builder.CreateLoad(Arg), Addr);
1216     Offset += Size;
1217     ++I;
1218   }
1219 
1220   FinishFunction();
1221 
1222   return Fn;
1223 }
1224 
1225 RValue CodeGenFunction::emitBuiltinOSLogFormat(const CallExpr &E) {
1226   assert(E.getNumArgs() >= 2 &&
1227          "__builtin_os_log_format takes at least 2 arguments");
1228   ASTContext &Ctx = getContext();
1229   analyze_os_log::OSLogBufferLayout Layout;
1230   analyze_os_log::computeOSLogBufferLayout(Ctx, &E, Layout);
1231   Address BufAddr = EmitPointerWithAlignment(E.getArg(0));
1232   llvm::SmallVector<llvm::Value *, 4> RetainableOperands;
1233 
1234   // Ignore argument 1, the format string. It is not currently used.
1235   CallArgList Args;
1236   Args.add(RValue::get(BufAddr.getPointer()), Ctx.VoidPtrTy);
1237 
1238   for (const auto &Item : Layout.Items) {
1239     int Size = Item.getSizeByte();
1240     if (!Size)
1241       continue;
1242 
1243     llvm::Value *ArgVal;
1244 
1245     if (Item.getKind() == analyze_os_log::OSLogBufferItem::MaskKind) {
1246       uint64_t Val = 0;
1247       for (unsigned I = 0, E = Item.getMaskType().size(); I < E; ++I)
1248         Val |= ((uint64_t)Item.getMaskType()[I]) << I * 8;
1249       ArgVal = llvm::Constant::getIntegerValue(Int64Ty, llvm::APInt(64, Val));
1250     } else if (const Expr *TheExpr = Item.getExpr()) {
1251       ArgVal = EmitScalarExpr(TheExpr, /*Ignore*/ false);
1252 
1253       // Check if this is a retainable type.
1254       if (TheExpr->getType()->isObjCRetainableType()) {
1255         assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar &&
1256                "Only scalar can be a ObjC retainable type");
1257         // Check if the object is constant, if not, save it in
1258         // RetainableOperands.
1259         if (!isa<Constant>(ArgVal))
1260           RetainableOperands.push_back(ArgVal);
1261       }
1262     } else {
1263       ArgVal = Builder.getInt32(Item.getConstValue().getQuantity());
1264     }
1265 
1266     unsigned ArgValSize =
1267         CGM.getDataLayout().getTypeSizeInBits(ArgVal->getType());
1268     llvm::IntegerType *IntTy = llvm::Type::getIntNTy(getLLVMContext(),
1269                                                      ArgValSize);
1270     ArgVal = Builder.CreateBitOrPointerCast(ArgVal, IntTy);
1271     CanQualType ArgTy = getOSLogArgType(Ctx, Size);
1272     // If ArgVal has type x86_fp80, zero-extend ArgVal.
1273     ArgVal = Builder.CreateZExtOrBitCast(ArgVal, ConvertType(ArgTy));
1274     Args.add(RValue::get(ArgVal), ArgTy);
1275   }
1276 
1277   const CGFunctionInfo &FI =
1278       CGM.getTypes().arrangeBuiltinFunctionCall(Ctx.VoidTy, Args);
1279   llvm::Function *F = CodeGenFunction(CGM).generateBuiltinOSLogHelperFunction(
1280       Layout, BufAddr.getAlignment());
1281   EmitCall(FI, CGCallee::forDirect(F), ReturnValueSlot(), Args);
1282 
1283   // Push a clang.arc.use cleanup for each object in RetainableOperands. The
1284   // cleanup will cause the use to appear after the final log call, keeping
1285   // the object valid while it’s held in the log buffer.  Note that if there’s
1286   // a release cleanup on the object, it will already be active; since
1287   // cleanups are emitted in reverse order, the use will occur before the
1288   // object is released.
1289   if (!RetainableOperands.empty() && getLangOpts().ObjCAutoRefCount &&
1290       CGM.getCodeGenOpts().OptimizationLevel != 0)
1291     for (llvm::Value *Object : RetainableOperands)
1292       pushFullExprCleanup<CallObjCArcUse>(getARCCleanupKind(), Object);
1293 
1294   return RValue::get(BufAddr.getPointer());
1295 }
1296 
1297 /// Determine if a binop is a checked mixed-sign multiply we can specialize.
1298 static bool isSpecialMixedSignMultiply(unsigned BuiltinID,
1299                                        WidthAndSignedness Op1Info,
1300                                        WidthAndSignedness Op2Info,
1301                                        WidthAndSignedness ResultInfo) {
1302   return BuiltinID == Builtin::BI__builtin_mul_overflow &&
1303          std::max(Op1Info.Width, Op2Info.Width) >= ResultInfo.Width &&
1304          Op1Info.Signed != Op2Info.Signed;
1305 }
1306 
1307 /// Emit a checked mixed-sign multiply. This is a cheaper specialization of
1308 /// the generic checked-binop irgen.
1309 static RValue
1310 EmitCheckedMixedSignMultiply(CodeGenFunction &CGF, const clang::Expr *Op1,
1311                              WidthAndSignedness Op1Info, const clang::Expr *Op2,
1312                              WidthAndSignedness Op2Info,
1313                              const clang::Expr *ResultArg, QualType ResultQTy,
1314                              WidthAndSignedness ResultInfo) {
1315   assert(isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info,
1316                                     Op2Info, ResultInfo) &&
1317          "Not a mixed-sign multipliction we can specialize");
1318 
1319   // Emit the signed and unsigned operands.
1320   const clang::Expr *SignedOp = Op1Info.Signed ? Op1 : Op2;
1321   const clang::Expr *UnsignedOp = Op1Info.Signed ? Op2 : Op1;
1322   llvm::Value *Signed = CGF.EmitScalarExpr(SignedOp);
1323   llvm::Value *Unsigned = CGF.EmitScalarExpr(UnsignedOp);
1324   unsigned SignedOpWidth = Op1Info.Signed ? Op1Info.Width : Op2Info.Width;
1325   unsigned UnsignedOpWidth = Op1Info.Signed ? Op2Info.Width : Op1Info.Width;
1326 
1327   // One of the operands may be smaller than the other. If so, [s|z]ext it.
1328   if (SignedOpWidth < UnsignedOpWidth)
1329     Signed = CGF.Builder.CreateSExt(Signed, Unsigned->getType(), "op.sext");
1330   if (UnsignedOpWidth < SignedOpWidth)
1331     Unsigned = CGF.Builder.CreateZExt(Unsigned, Signed->getType(), "op.zext");
1332 
1333   llvm::Type *OpTy = Signed->getType();
1334   llvm::Value *Zero = llvm::Constant::getNullValue(OpTy);
1335   Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg);
1336   llvm::Type *ResTy = ResultPtr.getElementType();
1337   unsigned OpWidth = std::max(Op1Info.Width, Op2Info.Width);
1338 
1339   // Take the absolute value of the signed operand.
1340   llvm::Value *IsNegative = CGF.Builder.CreateICmpSLT(Signed, Zero);
1341   llvm::Value *AbsOfNegative = CGF.Builder.CreateSub(Zero, Signed);
1342   llvm::Value *AbsSigned =
1343       CGF.Builder.CreateSelect(IsNegative, AbsOfNegative, Signed);
1344 
1345   // Perform a checked unsigned multiplication.
1346   llvm::Value *UnsignedOverflow;
1347   llvm::Value *UnsignedResult =
1348       EmitOverflowIntrinsic(CGF, llvm::Intrinsic::umul_with_overflow, AbsSigned,
1349                             Unsigned, UnsignedOverflow);
1350 
1351   llvm::Value *Overflow, *Result;
1352   if (ResultInfo.Signed) {
1353     // Signed overflow occurs if the result is greater than INT_MAX or lesser
1354     // than INT_MIN, i.e when |Result| > (INT_MAX + IsNegative).
1355     auto IntMax =
1356         llvm::APInt::getSignedMaxValue(ResultInfo.Width).zextOrSelf(OpWidth);
1357     llvm::Value *MaxResult =
1358         CGF.Builder.CreateAdd(llvm::ConstantInt::get(OpTy, IntMax),
1359                               CGF.Builder.CreateZExt(IsNegative, OpTy));
1360     llvm::Value *SignedOverflow =
1361         CGF.Builder.CreateICmpUGT(UnsignedResult, MaxResult);
1362     Overflow = CGF.Builder.CreateOr(UnsignedOverflow, SignedOverflow);
1363 
1364     // Prepare the signed result (possibly by negating it).
1365     llvm::Value *NegativeResult = CGF.Builder.CreateNeg(UnsignedResult);
1366     llvm::Value *SignedResult =
1367         CGF.Builder.CreateSelect(IsNegative, NegativeResult, UnsignedResult);
1368     Result = CGF.Builder.CreateTrunc(SignedResult, ResTy);
1369   } else {
1370     // Unsigned overflow occurs if the result is < 0 or greater than UINT_MAX.
1371     llvm::Value *Underflow = CGF.Builder.CreateAnd(
1372         IsNegative, CGF.Builder.CreateIsNotNull(UnsignedResult));
1373     Overflow = CGF.Builder.CreateOr(UnsignedOverflow, Underflow);
1374     if (ResultInfo.Width < OpWidth) {
1375       auto IntMax =
1376           llvm::APInt::getMaxValue(ResultInfo.Width).zext(OpWidth);
1377       llvm::Value *TruncOverflow = CGF.Builder.CreateICmpUGT(
1378           UnsignedResult, llvm::ConstantInt::get(OpTy, IntMax));
1379       Overflow = CGF.Builder.CreateOr(Overflow, TruncOverflow);
1380     }
1381 
1382     // Negate the product if it would be negative in infinite precision.
1383     Result = CGF.Builder.CreateSelect(
1384         IsNegative, CGF.Builder.CreateNeg(UnsignedResult), UnsignedResult);
1385 
1386     Result = CGF.Builder.CreateTrunc(Result, ResTy);
1387   }
1388   assert(Overflow && Result && "Missing overflow or result");
1389 
1390   bool isVolatile =
1391       ResultArg->getType()->getPointeeType().isVolatileQualified();
1392   CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr,
1393                           isVolatile);
1394   return RValue::get(Overflow);
1395 }
1396 
1397 static llvm::Value *dumpRecord(CodeGenFunction &CGF, QualType RType,
1398                                Value *&RecordPtr, CharUnits Align,
1399                                llvm::FunctionCallee Func, int Lvl) {
1400   ASTContext &Context = CGF.getContext();
1401   RecordDecl *RD = RType->castAs<RecordType>()->getDecl()->getDefinition();
1402   std::string Pad = std::string(Lvl * 4, ' ');
1403 
1404   Value *GString =
1405       CGF.Builder.CreateGlobalStringPtr(RType.getAsString() + " {\n");
1406   Value *Res = CGF.Builder.CreateCall(Func, {GString});
1407 
1408   static llvm::DenseMap<QualType, const char *> Types;
1409   if (Types.empty()) {
1410     Types[Context.CharTy] = "%c";
1411     Types[Context.BoolTy] = "%d";
1412     Types[Context.SignedCharTy] = "%hhd";
1413     Types[Context.UnsignedCharTy] = "%hhu";
1414     Types[Context.IntTy] = "%d";
1415     Types[Context.UnsignedIntTy] = "%u";
1416     Types[Context.LongTy] = "%ld";
1417     Types[Context.UnsignedLongTy] = "%lu";
1418     Types[Context.LongLongTy] = "%lld";
1419     Types[Context.UnsignedLongLongTy] = "%llu";
1420     Types[Context.ShortTy] = "%hd";
1421     Types[Context.UnsignedShortTy] = "%hu";
1422     Types[Context.VoidPtrTy] = "%p";
1423     Types[Context.FloatTy] = "%f";
1424     Types[Context.DoubleTy] = "%f";
1425     Types[Context.LongDoubleTy] = "%Lf";
1426     Types[Context.getPointerType(Context.CharTy)] = "%s";
1427     Types[Context.getPointerType(Context.getConstType(Context.CharTy))] = "%s";
1428   }
1429 
1430   for (const auto *FD : RD->fields()) {
1431     Value *FieldPtr = RecordPtr;
1432     if (RD->isUnion())
1433       FieldPtr = CGF.Builder.CreatePointerCast(
1434           FieldPtr, CGF.ConvertType(Context.getPointerType(FD->getType())));
1435     else
1436       FieldPtr = CGF.Builder.CreateStructGEP(CGF.ConvertType(RType), FieldPtr,
1437                                              FD->getFieldIndex());
1438 
1439     GString = CGF.Builder.CreateGlobalStringPtr(
1440         llvm::Twine(Pad)
1441             .concat(FD->getType().getAsString())
1442             .concat(llvm::Twine(' '))
1443             .concat(FD->getNameAsString())
1444             .concat(" : ")
1445             .str());
1446     Value *TmpRes = CGF.Builder.CreateCall(Func, {GString});
1447     Res = CGF.Builder.CreateAdd(Res, TmpRes);
1448 
1449     QualType CanonicalType =
1450         FD->getType().getUnqualifiedType().getCanonicalType();
1451 
1452     // We check whether we are in a recursive type
1453     if (CanonicalType->isRecordType()) {
1454       Value *TmpRes =
1455           dumpRecord(CGF, CanonicalType, FieldPtr, Align, Func, Lvl + 1);
1456       Res = CGF.Builder.CreateAdd(TmpRes, Res);
1457       continue;
1458     }
1459 
1460     // We try to determine the best format to print the current field
1461     llvm::Twine Format = Types.find(CanonicalType) == Types.end()
1462                              ? Types[Context.VoidPtrTy]
1463                              : Types[CanonicalType];
1464 
1465     Address FieldAddress = Address(FieldPtr, Align);
1466     FieldPtr = CGF.Builder.CreateLoad(FieldAddress);
1467 
1468     // FIXME Need to handle bitfield here
1469     GString = CGF.Builder.CreateGlobalStringPtr(
1470         Format.concat(llvm::Twine('\n')).str());
1471     TmpRes = CGF.Builder.CreateCall(Func, {GString, FieldPtr});
1472     Res = CGF.Builder.CreateAdd(Res, TmpRes);
1473   }
1474 
1475   GString = CGF.Builder.CreateGlobalStringPtr(Pad + "}\n");
1476   Value *TmpRes = CGF.Builder.CreateCall(Func, {GString});
1477   Res = CGF.Builder.CreateAdd(Res, TmpRes);
1478   return Res;
1479 }
1480 
1481 static bool
1482 TypeRequiresBuiltinLaunderImp(const ASTContext &Ctx, QualType Ty,
1483                               llvm::SmallPtrSetImpl<const Decl *> &Seen) {
1484   if (const auto *Arr = Ctx.getAsArrayType(Ty))
1485     Ty = Ctx.getBaseElementType(Arr);
1486 
1487   const auto *Record = Ty->getAsCXXRecordDecl();
1488   if (!Record)
1489     return false;
1490 
1491   // We've already checked this type, or are in the process of checking it.
1492   if (!Seen.insert(Record).second)
1493     return false;
1494 
1495   assert(Record->hasDefinition() &&
1496          "Incomplete types should already be diagnosed");
1497 
1498   if (Record->isDynamicClass())
1499     return true;
1500 
1501   for (FieldDecl *F : Record->fields()) {
1502     if (TypeRequiresBuiltinLaunderImp(Ctx, F->getType(), Seen))
1503       return true;
1504   }
1505   return false;
1506 }
1507 
1508 /// Determine if the specified type requires laundering by checking if it is a
1509 /// dynamic class type or contains a subobject which is a dynamic class type.
1510 static bool TypeRequiresBuiltinLaunder(CodeGenModule &CGM, QualType Ty) {
1511   if (!CGM.getCodeGenOpts().StrictVTablePointers)
1512     return false;
1513   llvm::SmallPtrSet<const Decl *, 16> Seen;
1514   return TypeRequiresBuiltinLaunderImp(CGM.getContext(), Ty, Seen);
1515 }
1516 
1517 RValue CodeGenFunction::emitRotate(const CallExpr *E, bool IsRotateRight) {
1518   llvm::Value *Src = EmitScalarExpr(E->getArg(0));
1519   llvm::Value *ShiftAmt = EmitScalarExpr(E->getArg(1));
1520 
1521   // The builtin's shift arg may have a different type than the source arg and
1522   // result, but the LLVM intrinsic uses the same type for all values.
1523   llvm::Type *Ty = Src->getType();
1524   ShiftAmt = Builder.CreateIntCast(ShiftAmt, Ty, false);
1525 
1526   // Rotate is a special case of LLVM funnel shift - 1st 2 args are the same.
1527   unsigned IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl;
1528   Function *F = CGM.getIntrinsic(IID, Ty);
1529   return RValue::get(Builder.CreateCall(F, { Src, Src, ShiftAmt }));
1530 }
1531 
1532 RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
1533                                         const CallExpr *E,
1534                                         ReturnValueSlot ReturnValue) {
1535   const FunctionDecl *FD = GD.getDecl()->getAsFunction();
1536   // See if we can constant fold this builtin.  If so, don't emit it at all.
1537   Expr::EvalResult Result;
1538   if (E->EvaluateAsRValue(Result, CGM.getContext()) &&
1539       !Result.hasSideEffects()) {
1540     if (Result.Val.isInt())
1541       return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
1542                                                 Result.Val.getInt()));
1543     if (Result.Val.isFloat())
1544       return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
1545                                                Result.Val.getFloat()));
1546   }
1547 
1548   // There are LLVM math intrinsics/instructions corresponding to math library
1549   // functions except the LLVM op will never set errno while the math library
1550   // might. Also, math builtins have the same semantics as their math library
1551   // twins. Thus, we can transform math library and builtin calls to their
1552   // LLVM counterparts if the call is marked 'const' (known to never set errno).
1553   if (FD->hasAttr<ConstAttr>()) {
1554     switch (BuiltinID) {
1555     case Builtin::BIceil:
1556     case Builtin::BIceilf:
1557     case Builtin::BIceill:
1558     case Builtin::BI__builtin_ceil:
1559     case Builtin::BI__builtin_ceilf:
1560     case Builtin::BI__builtin_ceilf16:
1561     case Builtin::BI__builtin_ceill:
1562       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::ceil));
1563 
1564     case Builtin::BIcopysign:
1565     case Builtin::BIcopysignf:
1566     case Builtin::BIcopysignl:
1567     case Builtin::BI__builtin_copysign:
1568     case Builtin::BI__builtin_copysignf:
1569     case Builtin::BI__builtin_copysignf16:
1570     case Builtin::BI__builtin_copysignl:
1571     case Builtin::BI__builtin_copysignf128:
1572       return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::copysign));
1573 
1574     case Builtin::BIcos:
1575     case Builtin::BIcosf:
1576     case Builtin::BIcosl:
1577     case Builtin::BI__builtin_cos:
1578     case Builtin::BI__builtin_cosf:
1579     case Builtin::BI__builtin_cosf16:
1580     case Builtin::BI__builtin_cosl:
1581       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::cos));
1582 
1583     case Builtin::BIexp:
1584     case Builtin::BIexpf:
1585     case Builtin::BIexpl:
1586     case Builtin::BI__builtin_exp:
1587     case Builtin::BI__builtin_expf:
1588     case Builtin::BI__builtin_expf16:
1589     case Builtin::BI__builtin_expl:
1590       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::exp));
1591 
1592     case Builtin::BIexp2:
1593     case Builtin::BIexp2f:
1594     case Builtin::BIexp2l:
1595     case Builtin::BI__builtin_exp2:
1596     case Builtin::BI__builtin_exp2f:
1597     case Builtin::BI__builtin_exp2f16:
1598     case Builtin::BI__builtin_exp2l:
1599       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::exp2));
1600 
1601     case Builtin::BIfabs:
1602     case Builtin::BIfabsf:
1603     case Builtin::BIfabsl:
1604     case Builtin::BI__builtin_fabs:
1605     case Builtin::BI__builtin_fabsf:
1606     case Builtin::BI__builtin_fabsf16:
1607     case Builtin::BI__builtin_fabsl:
1608     case Builtin::BI__builtin_fabsf128:
1609       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs));
1610 
1611     case Builtin::BIfloor:
1612     case Builtin::BIfloorf:
1613     case Builtin::BIfloorl:
1614     case Builtin::BI__builtin_floor:
1615     case Builtin::BI__builtin_floorf:
1616     case Builtin::BI__builtin_floorf16:
1617     case Builtin::BI__builtin_floorl:
1618       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::floor));
1619 
1620     case Builtin::BIfma:
1621     case Builtin::BIfmaf:
1622     case Builtin::BIfmal:
1623     case Builtin::BI__builtin_fma:
1624     case Builtin::BI__builtin_fmaf:
1625     case Builtin::BI__builtin_fmaf16:
1626     case Builtin::BI__builtin_fmal:
1627       return RValue::get(emitTernaryBuiltin(*this, E, Intrinsic::fma));
1628 
1629     case Builtin::BIfmax:
1630     case Builtin::BIfmaxf:
1631     case Builtin::BIfmaxl:
1632     case Builtin::BI__builtin_fmax:
1633     case Builtin::BI__builtin_fmaxf:
1634     case Builtin::BI__builtin_fmaxf16:
1635     case Builtin::BI__builtin_fmaxl:
1636       return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::maxnum));
1637 
1638     case Builtin::BIfmin:
1639     case Builtin::BIfminf:
1640     case Builtin::BIfminl:
1641     case Builtin::BI__builtin_fmin:
1642     case Builtin::BI__builtin_fminf:
1643     case Builtin::BI__builtin_fminf16:
1644     case Builtin::BI__builtin_fminl:
1645       return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::minnum));
1646 
1647     // fmod() is a special-case. It maps to the frem instruction rather than an
1648     // LLVM intrinsic.
1649     case Builtin::BIfmod:
1650     case Builtin::BIfmodf:
1651     case Builtin::BIfmodl:
1652     case Builtin::BI__builtin_fmod:
1653     case Builtin::BI__builtin_fmodf:
1654     case Builtin::BI__builtin_fmodf16:
1655     case Builtin::BI__builtin_fmodl: {
1656       Value *Arg1 = EmitScalarExpr(E->getArg(0));
1657       Value *Arg2 = EmitScalarExpr(E->getArg(1));
1658       return RValue::get(Builder.CreateFRem(Arg1, Arg2, "fmod"));
1659     }
1660 
1661     case Builtin::BIlog:
1662     case Builtin::BIlogf:
1663     case Builtin::BIlogl:
1664     case Builtin::BI__builtin_log:
1665     case Builtin::BI__builtin_logf:
1666     case Builtin::BI__builtin_logf16:
1667     case Builtin::BI__builtin_logl:
1668       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log));
1669 
1670     case Builtin::BIlog10:
1671     case Builtin::BIlog10f:
1672     case Builtin::BIlog10l:
1673     case Builtin::BI__builtin_log10:
1674     case Builtin::BI__builtin_log10f:
1675     case Builtin::BI__builtin_log10f16:
1676     case Builtin::BI__builtin_log10l:
1677       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log10));
1678 
1679     case Builtin::BIlog2:
1680     case Builtin::BIlog2f:
1681     case Builtin::BIlog2l:
1682     case Builtin::BI__builtin_log2:
1683     case Builtin::BI__builtin_log2f:
1684     case Builtin::BI__builtin_log2f16:
1685     case Builtin::BI__builtin_log2l:
1686       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log2));
1687 
1688     case Builtin::BInearbyint:
1689     case Builtin::BInearbyintf:
1690     case Builtin::BInearbyintl:
1691     case Builtin::BI__builtin_nearbyint:
1692     case Builtin::BI__builtin_nearbyintf:
1693     case Builtin::BI__builtin_nearbyintl:
1694       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::nearbyint));
1695 
1696     case Builtin::BIpow:
1697     case Builtin::BIpowf:
1698     case Builtin::BIpowl:
1699     case Builtin::BI__builtin_pow:
1700     case Builtin::BI__builtin_powf:
1701     case Builtin::BI__builtin_powf16:
1702     case Builtin::BI__builtin_powl:
1703       return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::pow));
1704 
1705     case Builtin::BIrint:
1706     case Builtin::BIrintf:
1707     case Builtin::BIrintl:
1708     case Builtin::BI__builtin_rint:
1709     case Builtin::BI__builtin_rintf:
1710     case Builtin::BI__builtin_rintf16:
1711     case Builtin::BI__builtin_rintl:
1712       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::rint));
1713 
1714     case Builtin::BIround:
1715     case Builtin::BIroundf:
1716     case Builtin::BIroundl:
1717     case Builtin::BI__builtin_round:
1718     case Builtin::BI__builtin_roundf:
1719     case Builtin::BI__builtin_roundf16:
1720     case Builtin::BI__builtin_roundl:
1721       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::round));
1722 
1723     case Builtin::BIsin:
1724     case Builtin::BIsinf:
1725     case Builtin::BIsinl:
1726     case Builtin::BI__builtin_sin:
1727     case Builtin::BI__builtin_sinf:
1728     case Builtin::BI__builtin_sinf16:
1729     case Builtin::BI__builtin_sinl:
1730       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::sin));
1731 
1732     case Builtin::BIsqrt:
1733     case Builtin::BIsqrtf:
1734     case Builtin::BIsqrtl:
1735     case Builtin::BI__builtin_sqrt:
1736     case Builtin::BI__builtin_sqrtf:
1737     case Builtin::BI__builtin_sqrtf16:
1738     case Builtin::BI__builtin_sqrtl:
1739       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::sqrt));
1740 
1741     case Builtin::BItrunc:
1742     case Builtin::BItruncf:
1743     case Builtin::BItruncl:
1744     case Builtin::BI__builtin_trunc:
1745     case Builtin::BI__builtin_truncf:
1746     case Builtin::BI__builtin_truncf16:
1747     case Builtin::BI__builtin_truncl:
1748       return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::trunc));
1749 
1750     case Builtin::BIlround:
1751     case Builtin::BIlroundf:
1752     case Builtin::BIlroundl:
1753     case Builtin::BI__builtin_lround:
1754     case Builtin::BI__builtin_lroundf:
1755     case Builtin::BI__builtin_lroundl:
1756       return RValue::get(emitFPToIntRoundBuiltin(*this, E, Intrinsic::lround));
1757 
1758     case Builtin::BIllround:
1759     case Builtin::BIllroundf:
1760     case Builtin::BIllroundl:
1761     case Builtin::BI__builtin_llround:
1762     case Builtin::BI__builtin_llroundf:
1763     case Builtin::BI__builtin_llroundl:
1764       return RValue::get(emitFPToIntRoundBuiltin(*this, E, Intrinsic::llround));
1765 
1766     case Builtin::BIlrint:
1767     case Builtin::BIlrintf:
1768     case Builtin::BIlrintl:
1769     case Builtin::BI__builtin_lrint:
1770     case Builtin::BI__builtin_lrintf:
1771     case Builtin::BI__builtin_lrintl:
1772       return RValue::get(emitFPToIntRoundBuiltin(*this, E, Intrinsic::lrint));
1773 
1774     case Builtin::BIllrint:
1775     case Builtin::BIllrintf:
1776     case Builtin::BIllrintl:
1777     case Builtin::BI__builtin_llrint:
1778     case Builtin::BI__builtin_llrintf:
1779     case Builtin::BI__builtin_llrintl:
1780       return RValue::get(emitFPToIntRoundBuiltin(*this, E, Intrinsic::llrint));
1781 
1782     default:
1783       break;
1784     }
1785   }
1786 
1787   switch (BuiltinID) {
1788   default: break;
1789   case Builtin::BI__builtin___CFStringMakeConstantString:
1790   case Builtin::BI__builtin___NSStringMakeConstantString:
1791     return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType()));
1792   case Builtin::BI__builtin_stdarg_start:
1793   case Builtin::BI__builtin_va_start:
1794   case Builtin::BI__va_start:
1795   case Builtin::BI__builtin_va_end:
1796     return RValue::get(
1797         EmitVAStartEnd(BuiltinID == Builtin::BI__va_start
1798                            ? EmitScalarExpr(E->getArg(0))
1799                            : EmitVAListRef(E->getArg(0)).getPointer(),
1800                        BuiltinID != Builtin::BI__builtin_va_end));
1801   case Builtin::BI__builtin_va_copy: {
1802     Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer();
1803     Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer();
1804 
1805     llvm::Type *Type = Int8PtrTy;
1806 
1807     DstPtr = Builder.CreateBitCast(DstPtr, Type);
1808     SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
1809     return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy),
1810                                           {DstPtr, SrcPtr}));
1811   }
1812   case Builtin::BI__builtin_abs:
1813   case Builtin::BI__builtin_labs:
1814   case Builtin::BI__builtin_llabs: {
1815     // X < 0 ? -X : X
1816     // The negation has 'nsw' because abs of INT_MIN is undefined.
1817     Value *ArgValue = EmitScalarExpr(E->getArg(0));
1818     Value *NegOp = Builder.CreateNSWNeg(ArgValue, "neg");
1819     Constant *Zero = llvm::Constant::getNullValue(ArgValue->getType());
1820     Value *CmpResult = Builder.CreateICmpSLT(ArgValue, Zero, "abscond");
1821     Value *Result = Builder.CreateSelect(CmpResult, NegOp, ArgValue, "abs");
1822     return RValue::get(Result);
1823   }
1824   case Builtin::BI__builtin_conj:
1825   case Builtin::BI__builtin_conjf:
1826   case Builtin::BI__builtin_conjl: {
1827     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
1828     Value *Real = ComplexVal.first;
1829     Value *Imag = ComplexVal.second;
1830     Value *Zero =
1831       Imag->getType()->isFPOrFPVectorTy()
1832         ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType())
1833         : llvm::Constant::getNullValue(Imag->getType());
1834 
1835     Imag = Builder.CreateFSub(Zero, Imag, "sub");
1836     return RValue::getComplex(std::make_pair(Real, Imag));
1837   }
1838   case Builtin::BI__builtin_creal:
1839   case Builtin::BI__builtin_crealf:
1840   case Builtin::BI__builtin_creall:
1841   case Builtin::BIcreal:
1842   case Builtin::BIcrealf:
1843   case Builtin::BIcreall: {
1844     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
1845     return RValue::get(ComplexVal.first);
1846   }
1847 
1848   case Builtin::BI__builtin_dump_struct: {
1849     llvm::Type *LLVMIntTy = getTypes().ConvertType(getContext().IntTy);
1850     llvm::FunctionType *LLVMFuncType = llvm::FunctionType::get(
1851         LLVMIntTy, {llvm::Type::getInt8PtrTy(getLLVMContext())}, true);
1852 
1853     Value *Func = EmitScalarExpr(E->getArg(1)->IgnoreImpCasts());
1854     CharUnits Arg0Align = EmitPointerWithAlignment(E->getArg(0)).getAlignment();
1855 
1856     const Expr *Arg0 = E->getArg(0)->IgnoreImpCasts();
1857     QualType Arg0Type = Arg0->getType()->getPointeeType();
1858 
1859     Value *RecordPtr = EmitScalarExpr(Arg0);
1860     Value *Res = dumpRecord(*this, Arg0Type, RecordPtr, Arg0Align,
1861                             {LLVMFuncType, Func}, 0);
1862     return RValue::get(Res);
1863   }
1864 
1865   case Builtin::BI__builtin_preserve_access_index: {
1866     // Only enabled preserved access index region when debuginfo
1867     // is available as debuginfo is needed to preserve user-level
1868     // access pattern.
1869     if (!getDebugInfo()) {
1870       CGM.Error(E->getExprLoc(), "using builtin_preserve_access_index() without -g");
1871       return RValue::get(EmitScalarExpr(E->getArg(0)));
1872     }
1873 
1874     // Nested builtin_preserve_access_index() not supported
1875     if (IsInPreservedAIRegion) {
1876       CGM.Error(E->getExprLoc(), "nested builtin_preserve_access_index() not supported");
1877       return RValue::get(EmitScalarExpr(E->getArg(0)));
1878     }
1879 
1880     IsInPreservedAIRegion = true;
1881     Value *Res = EmitScalarExpr(E->getArg(0));
1882     IsInPreservedAIRegion = false;
1883     return RValue::get(Res);
1884   }
1885 
1886   case Builtin::BI__builtin_cimag:
1887   case Builtin::BI__builtin_cimagf:
1888   case Builtin::BI__builtin_cimagl:
1889   case Builtin::BIcimag:
1890   case Builtin::BIcimagf:
1891   case Builtin::BIcimagl: {
1892     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
1893     return RValue::get(ComplexVal.second);
1894   }
1895 
1896   case Builtin::BI__builtin_clrsb:
1897   case Builtin::BI__builtin_clrsbl:
1898   case Builtin::BI__builtin_clrsbll: {
1899     // clrsb(x) -> clz(x < 0 ? ~x : x) - 1 or
1900     Value *ArgValue = EmitScalarExpr(E->getArg(0));
1901 
1902     llvm::Type *ArgType = ArgValue->getType();
1903     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1904 
1905     llvm::Type *ResultType = ConvertType(E->getType());
1906     Value *Zero = llvm::Constant::getNullValue(ArgType);
1907     Value *IsNeg = Builder.CreateICmpSLT(ArgValue, Zero, "isneg");
1908     Value *Inverse = Builder.CreateNot(ArgValue, "not");
1909     Value *Tmp = Builder.CreateSelect(IsNeg, Inverse, ArgValue);
1910     Value *Ctlz = Builder.CreateCall(F, {Tmp, Builder.getFalse()});
1911     Value *Result = Builder.CreateSub(Ctlz, llvm::ConstantInt::get(ArgType, 1));
1912     Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1913                                    "cast");
1914     return RValue::get(Result);
1915   }
1916   case Builtin::BI__builtin_ctzs:
1917   case Builtin::BI__builtin_ctz:
1918   case Builtin::BI__builtin_ctzl:
1919   case Builtin::BI__builtin_ctzll: {
1920     Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero);
1921 
1922     llvm::Type *ArgType = ArgValue->getType();
1923     Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
1924 
1925     llvm::Type *ResultType = ConvertType(E->getType());
1926     Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
1927     Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
1928     if (Result->getType() != ResultType)
1929       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1930                                      "cast");
1931     return RValue::get(Result);
1932   }
1933   case Builtin::BI__builtin_clzs:
1934   case Builtin::BI__builtin_clz:
1935   case Builtin::BI__builtin_clzl:
1936   case Builtin::BI__builtin_clzll: {
1937     Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero);
1938 
1939     llvm::Type *ArgType = ArgValue->getType();
1940     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1941 
1942     llvm::Type *ResultType = ConvertType(E->getType());
1943     Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
1944     Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
1945     if (Result->getType() != ResultType)
1946       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1947                                      "cast");
1948     return RValue::get(Result);
1949   }
1950   case Builtin::BI__builtin_ffs:
1951   case Builtin::BI__builtin_ffsl:
1952   case Builtin::BI__builtin_ffsll: {
1953     // ffs(x) -> x ? cttz(x) + 1 : 0
1954     Value *ArgValue = EmitScalarExpr(E->getArg(0));
1955 
1956     llvm::Type *ArgType = ArgValue->getType();
1957     Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
1958 
1959     llvm::Type *ResultType = ConvertType(E->getType());
1960     Value *Tmp =
1961         Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}),
1962                           llvm::ConstantInt::get(ArgType, 1));
1963     Value *Zero = llvm::Constant::getNullValue(ArgType);
1964     Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
1965     Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
1966     if (Result->getType() != ResultType)
1967       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1968                                      "cast");
1969     return RValue::get(Result);
1970   }
1971   case Builtin::BI__builtin_parity:
1972   case Builtin::BI__builtin_parityl:
1973   case Builtin::BI__builtin_parityll: {
1974     // parity(x) -> ctpop(x) & 1
1975     Value *ArgValue = EmitScalarExpr(E->getArg(0));
1976 
1977     llvm::Type *ArgType = ArgValue->getType();
1978     Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
1979 
1980     llvm::Type *ResultType = ConvertType(E->getType());
1981     Value *Tmp = Builder.CreateCall(F, ArgValue);
1982     Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
1983     if (Result->getType() != ResultType)
1984       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1985                                      "cast");
1986     return RValue::get(Result);
1987   }
1988   case Builtin::BI__lzcnt16:
1989   case Builtin::BI__lzcnt:
1990   case Builtin::BI__lzcnt64: {
1991     Value *ArgValue = EmitScalarExpr(E->getArg(0));
1992 
1993     llvm::Type *ArgType = ArgValue->getType();
1994     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1995 
1996     llvm::Type *ResultType = ConvertType(E->getType());
1997     Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getFalse()});
1998     if (Result->getType() != ResultType)
1999       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2000                                      "cast");
2001     return RValue::get(Result);
2002   }
2003   case Builtin::BI__popcnt16:
2004   case Builtin::BI__popcnt:
2005   case Builtin::BI__popcnt64:
2006   case Builtin::BI__builtin_popcount:
2007   case Builtin::BI__builtin_popcountl:
2008   case Builtin::BI__builtin_popcountll: {
2009     Value *ArgValue = EmitScalarExpr(E->getArg(0));
2010 
2011     llvm::Type *ArgType = ArgValue->getType();
2012     Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
2013 
2014     llvm::Type *ResultType = ConvertType(E->getType());
2015     Value *Result = Builder.CreateCall(F, ArgValue);
2016     if (Result->getType() != ResultType)
2017       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2018                                      "cast");
2019     return RValue::get(Result);
2020   }
2021   case Builtin::BI__builtin_unpredictable: {
2022     // Always return the argument of __builtin_unpredictable. LLVM does not
2023     // handle this builtin. Metadata for this builtin should be added directly
2024     // to instructions such as branches or switches that use it.
2025     return RValue::get(EmitScalarExpr(E->getArg(0)));
2026   }
2027   case Builtin::BI__builtin_expect: {
2028     Value *ArgValue = EmitScalarExpr(E->getArg(0));
2029     llvm::Type *ArgType = ArgValue->getType();
2030 
2031     Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
2032     // Don't generate llvm.expect on -O0 as the backend won't use it for
2033     // anything.
2034     // Note, we still IRGen ExpectedValue because it could have side-effects.
2035     if (CGM.getCodeGenOpts().OptimizationLevel == 0)
2036       return RValue::get(ArgValue);
2037 
2038     Function *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
2039     Value *Result =
2040         Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval");
2041     return RValue::get(Result);
2042   }
2043   case Builtin::BI__builtin_assume_aligned: {
2044     const Expr *Ptr = E->getArg(0);
2045     Value *PtrValue = EmitScalarExpr(Ptr);
2046     Value *OffsetValue =
2047       (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr;
2048 
2049     Value *AlignmentValue = EmitScalarExpr(E->getArg(1));
2050     ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue);
2051     if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment))
2052       AlignmentCI = ConstantInt::get(AlignmentCI->getType(),
2053                                      llvm::Value::MaximumAlignment);
2054 
2055     EmitAlignmentAssumption(PtrValue, Ptr,
2056                             /*The expr loc is sufficient.*/ SourceLocation(),
2057                             AlignmentCI, OffsetValue);
2058     return RValue::get(PtrValue);
2059   }
2060   case Builtin::BI__assume:
2061   case Builtin::BI__builtin_assume: {
2062     if (E->getArg(0)->HasSideEffects(getContext()))
2063       return RValue::get(nullptr);
2064 
2065     Value *ArgValue = EmitScalarExpr(E->getArg(0));
2066     Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
2067     return RValue::get(Builder.CreateCall(FnAssume, ArgValue));
2068   }
2069   case Builtin::BI__builtin_bswap16:
2070   case Builtin::BI__builtin_bswap32:
2071   case Builtin::BI__builtin_bswap64: {
2072     return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap));
2073   }
2074   case Builtin::BI__builtin_bitreverse8:
2075   case Builtin::BI__builtin_bitreverse16:
2076   case Builtin::BI__builtin_bitreverse32:
2077   case Builtin::BI__builtin_bitreverse64: {
2078     return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse));
2079   }
2080   case Builtin::BI__builtin_rotateleft8:
2081   case Builtin::BI__builtin_rotateleft16:
2082   case Builtin::BI__builtin_rotateleft32:
2083   case Builtin::BI__builtin_rotateleft64:
2084   case Builtin::BI_rotl8: // Microsoft variants of rotate left
2085   case Builtin::BI_rotl16:
2086   case Builtin::BI_rotl:
2087   case Builtin::BI_lrotl:
2088   case Builtin::BI_rotl64:
2089     return emitRotate(E, false);
2090 
2091   case Builtin::BI__builtin_rotateright8:
2092   case Builtin::BI__builtin_rotateright16:
2093   case Builtin::BI__builtin_rotateright32:
2094   case Builtin::BI__builtin_rotateright64:
2095   case Builtin::BI_rotr8: // Microsoft variants of rotate right
2096   case Builtin::BI_rotr16:
2097   case Builtin::BI_rotr:
2098   case Builtin::BI_lrotr:
2099   case Builtin::BI_rotr64:
2100     return emitRotate(E, true);
2101 
2102   case Builtin::BI__builtin_constant_p: {
2103     llvm::Type *ResultType = ConvertType(E->getType());
2104 
2105     const Expr *Arg = E->getArg(0);
2106     QualType ArgType = Arg->getType();
2107     // FIXME: The allowance for Obj-C pointers and block pointers is historical
2108     // and likely a mistake.
2109     if (!ArgType->isIntegralOrEnumerationType() && !ArgType->isFloatingType() &&
2110         !ArgType->isObjCObjectPointerType() && !ArgType->isBlockPointerType())
2111       // Per the GCC documentation, only numeric constants are recognized after
2112       // inlining.
2113       return RValue::get(ConstantInt::get(ResultType, 0));
2114 
2115     if (Arg->HasSideEffects(getContext()))
2116       // The argument is unevaluated, so be conservative if it might have
2117       // side-effects.
2118       return RValue::get(ConstantInt::get(ResultType, 0));
2119 
2120     Value *ArgValue = EmitScalarExpr(Arg);
2121     if (ArgType->isObjCObjectPointerType()) {
2122       // Convert Objective-C objects to id because we cannot distinguish between
2123       // LLVM types for Obj-C classes as they are opaque.
2124       ArgType = CGM.getContext().getObjCIdType();
2125       ArgValue = Builder.CreateBitCast(ArgValue, ConvertType(ArgType));
2126     }
2127     Function *F =
2128         CGM.getIntrinsic(Intrinsic::is_constant, ConvertType(ArgType));
2129     Value *Result = Builder.CreateCall(F, ArgValue);
2130     if (Result->getType() != ResultType)
2131       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/false);
2132     return RValue::get(Result);
2133   }
2134   case Builtin::BI__builtin_dynamic_object_size:
2135   case Builtin::BI__builtin_object_size: {
2136     unsigned Type =
2137         E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue();
2138     auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType()));
2139 
2140     // We pass this builtin onto the optimizer so that it can figure out the
2141     // object size in more complex cases.
2142     bool IsDynamic = BuiltinID == Builtin::BI__builtin_dynamic_object_size;
2143     return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType,
2144                                              /*EmittedE=*/nullptr, IsDynamic));
2145   }
2146   case Builtin::BI__builtin_prefetch: {
2147     Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
2148     // FIXME: Technically these constants should of type 'int', yes?
2149     RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
2150       llvm::ConstantInt::get(Int32Ty, 0);
2151     Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
2152       llvm::ConstantInt::get(Int32Ty, 3);
2153     Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
2154     Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
2155     return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data}));
2156   }
2157   case Builtin::BI__builtin_readcyclecounter: {
2158     Function *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
2159     return RValue::get(Builder.CreateCall(F));
2160   }
2161   case Builtin::BI__builtin___clear_cache: {
2162     Value *Begin = EmitScalarExpr(E->getArg(0));
2163     Value *End = EmitScalarExpr(E->getArg(1));
2164     Function *F = CGM.getIntrinsic(Intrinsic::clear_cache);
2165     return RValue::get(Builder.CreateCall(F, {Begin, End}));
2166   }
2167   case Builtin::BI__builtin_trap:
2168     return RValue::get(EmitTrapCall(Intrinsic::trap));
2169   case Builtin::BI__debugbreak:
2170     return RValue::get(EmitTrapCall(Intrinsic::debugtrap));
2171   case Builtin::BI__builtin_unreachable: {
2172     EmitUnreachable(E->getExprLoc());
2173 
2174     // We do need to preserve an insertion point.
2175     EmitBlock(createBasicBlock("unreachable.cont"));
2176 
2177     return RValue::get(nullptr);
2178   }
2179 
2180   case Builtin::BI__builtin_powi:
2181   case Builtin::BI__builtin_powif:
2182   case Builtin::BI__builtin_powil: {
2183     Value *Base = EmitScalarExpr(E->getArg(0));
2184     Value *Exponent = EmitScalarExpr(E->getArg(1));
2185     llvm::Type *ArgType = Base->getType();
2186     Function *F = CGM.getIntrinsic(Intrinsic::powi, ArgType);
2187     return RValue::get(Builder.CreateCall(F, {Base, Exponent}));
2188   }
2189 
2190   case Builtin::BI__builtin_isgreater:
2191   case Builtin::BI__builtin_isgreaterequal:
2192   case Builtin::BI__builtin_isless:
2193   case Builtin::BI__builtin_islessequal:
2194   case Builtin::BI__builtin_islessgreater:
2195   case Builtin::BI__builtin_isunordered: {
2196     // Ordered comparisons: we know the arguments to these are matching scalar
2197     // floating point values.
2198     Value *LHS = EmitScalarExpr(E->getArg(0));
2199     Value *RHS = EmitScalarExpr(E->getArg(1));
2200 
2201     switch (BuiltinID) {
2202     default: llvm_unreachable("Unknown ordered comparison");
2203     case Builtin::BI__builtin_isgreater:
2204       LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
2205       break;
2206     case Builtin::BI__builtin_isgreaterequal:
2207       LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
2208       break;
2209     case Builtin::BI__builtin_isless:
2210       LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
2211       break;
2212     case Builtin::BI__builtin_islessequal:
2213       LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
2214       break;
2215     case Builtin::BI__builtin_islessgreater:
2216       LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
2217       break;
2218     case Builtin::BI__builtin_isunordered:
2219       LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
2220       break;
2221     }
2222     // ZExt bool to int type.
2223     return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
2224   }
2225   case Builtin::BI__builtin_isnan: {
2226     Value *V = EmitScalarExpr(E->getArg(0));
2227     V = Builder.CreateFCmpUNO(V, V, "cmp");
2228     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
2229   }
2230 
2231   case Builtin::BIfinite:
2232   case Builtin::BI__finite:
2233   case Builtin::BIfinitef:
2234   case Builtin::BI__finitef:
2235   case Builtin::BIfinitel:
2236   case Builtin::BI__finitel:
2237   case Builtin::BI__builtin_isinf:
2238   case Builtin::BI__builtin_isfinite: {
2239     // isinf(x)    --> fabs(x) == infinity
2240     // isfinite(x) --> fabs(x) != infinity
2241     // x != NaN via the ordered compare in either case.
2242     Value *V = EmitScalarExpr(E->getArg(0));
2243     Value *Fabs = EmitFAbs(*this, V);
2244     Constant *Infinity = ConstantFP::getInfinity(V->getType());
2245     CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf)
2246                                   ? CmpInst::FCMP_OEQ
2247                                   : CmpInst::FCMP_ONE;
2248     Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf");
2249     return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType())));
2250   }
2251 
2252   case Builtin::BI__builtin_isinf_sign: {
2253     // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0
2254     Value *Arg = EmitScalarExpr(E->getArg(0));
2255     Value *AbsArg = EmitFAbs(*this, Arg);
2256     Value *IsInf = Builder.CreateFCmpOEQ(
2257         AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf");
2258     Value *IsNeg = EmitSignBit(*this, Arg);
2259 
2260     llvm::Type *IntTy = ConvertType(E->getType());
2261     Value *Zero = Constant::getNullValue(IntTy);
2262     Value *One = ConstantInt::get(IntTy, 1);
2263     Value *NegativeOne = ConstantInt::get(IntTy, -1);
2264     Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One);
2265     Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero);
2266     return RValue::get(Result);
2267   }
2268 
2269   case Builtin::BI__builtin_isnormal: {
2270     // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
2271     Value *V = EmitScalarExpr(E->getArg(0));
2272     Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
2273 
2274     Value *Abs = EmitFAbs(*this, V);
2275     Value *IsLessThanInf =
2276       Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
2277     APFloat Smallest = APFloat::getSmallestNormalized(
2278                    getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
2279     Value *IsNormal =
2280       Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
2281                             "isnormal");
2282     V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
2283     V = Builder.CreateAnd(V, IsNormal, "and");
2284     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
2285   }
2286 
2287   case Builtin::BI__builtin_flt_rounds: {
2288     Function *F = CGM.getIntrinsic(Intrinsic::flt_rounds);
2289 
2290     llvm::Type *ResultType = ConvertType(E->getType());
2291     Value *Result = Builder.CreateCall(F);
2292     if (Result->getType() != ResultType)
2293       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2294                                      "cast");
2295     return RValue::get(Result);
2296   }
2297 
2298   case Builtin::BI__builtin_fpclassify: {
2299     Value *V = EmitScalarExpr(E->getArg(5));
2300     llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
2301 
2302     // Create Result
2303     BasicBlock *Begin = Builder.GetInsertBlock();
2304     BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
2305     Builder.SetInsertPoint(End);
2306     PHINode *Result =
2307       Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
2308                         "fpclassify_result");
2309 
2310     // if (V==0) return FP_ZERO
2311     Builder.SetInsertPoint(Begin);
2312     Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
2313                                           "iszero");
2314     Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
2315     BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
2316     Builder.CreateCondBr(IsZero, End, NotZero);
2317     Result->addIncoming(ZeroLiteral, Begin);
2318 
2319     // if (V != V) return FP_NAN
2320     Builder.SetInsertPoint(NotZero);
2321     Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
2322     Value *NanLiteral = EmitScalarExpr(E->getArg(0));
2323     BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
2324     Builder.CreateCondBr(IsNan, End, NotNan);
2325     Result->addIncoming(NanLiteral, NotZero);
2326 
2327     // if (fabs(V) == infinity) return FP_INFINITY
2328     Builder.SetInsertPoint(NotNan);
2329     Value *VAbs = EmitFAbs(*this, V);
2330     Value *IsInf =
2331       Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
2332                             "isinf");
2333     Value *InfLiteral = EmitScalarExpr(E->getArg(1));
2334     BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
2335     Builder.CreateCondBr(IsInf, End, NotInf);
2336     Result->addIncoming(InfLiteral, NotNan);
2337 
2338     // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
2339     Builder.SetInsertPoint(NotInf);
2340     APFloat Smallest = APFloat::getSmallestNormalized(
2341         getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
2342     Value *IsNormal =
2343       Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
2344                             "isnormal");
2345     Value *NormalResult =
2346       Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
2347                            EmitScalarExpr(E->getArg(3)));
2348     Builder.CreateBr(End);
2349     Result->addIncoming(NormalResult, NotInf);
2350 
2351     // return Result
2352     Builder.SetInsertPoint(End);
2353     return RValue::get(Result);
2354   }
2355 
2356   case Builtin::BIalloca:
2357   case Builtin::BI_alloca:
2358   case Builtin::BI__builtin_alloca: {
2359     Value *Size = EmitScalarExpr(E->getArg(0));
2360     const TargetInfo &TI = getContext().getTargetInfo();
2361     // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__.
2362     unsigned SuitableAlignmentInBytes =
2363         CGM.getContext()
2364             .toCharUnitsFromBits(TI.getSuitableAlign())
2365             .getQuantity();
2366     AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
2367     AI->setAlignment(MaybeAlign(SuitableAlignmentInBytes));
2368     initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes);
2369     return RValue::get(AI);
2370   }
2371 
2372   case Builtin::BI__builtin_alloca_with_align: {
2373     Value *Size = EmitScalarExpr(E->getArg(0));
2374     Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1));
2375     auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue);
2376     unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue();
2377     unsigned AlignmentInBytes =
2378         CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getQuantity();
2379     AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
2380     AI->setAlignment(MaybeAlign(AlignmentInBytes));
2381     initializeAlloca(*this, AI, Size, AlignmentInBytes);
2382     return RValue::get(AI);
2383   }
2384 
2385   case Builtin::BIbzero:
2386   case Builtin::BI__builtin_bzero: {
2387     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2388     Value *SizeVal = EmitScalarExpr(E->getArg(1));
2389     EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2390                         E->getArg(0)->getExprLoc(), FD, 0);
2391     Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false);
2392     return RValue::get(nullptr);
2393   }
2394   case Builtin::BImemcpy:
2395   case Builtin::BI__builtin_memcpy: {
2396     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2397     Address Src = EmitPointerWithAlignment(E->getArg(1));
2398     Value *SizeVal = EmitScalarExpr(E->getArg(2));
2399     EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2400                         E->getArg(0)->getExprLoc(), FD, 0);
2401     EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
2402                         E->getArg(1)->getExprLoc(), FD, 1);
2403     Builder.CreateMemCpy(Dest, Src, SizeVal, false);
2404     return RValue::get(Dest.getPointer());
2405   }
2406 
2407   case Builtin::BI__builtin_char_memchr:
2408     BuiltinID = Builtin::BI__builtin_memchr;
2409     break;
2410 
2411   case Builtin::BI__builtin___memcpy_chk: {
2412     // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
2413     Expr::EvalResult SizeResult, DstSizeResult;
2414     if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
2415         !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
2416       break;
2417     llvm::APSInt Size = SizeResult.Val.getInt();
2418     llvm::APSInt DstSize = DstSizeResult.Val.getInt();
2419     if (Size.ugt(DstSize))
2420       break;
2421     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2422     Address Src = EmitPointerWithAlignment(E->getArg(1));
2423     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
2424     Builder.CreateMemCpy(Dest, Src, SizeVal, false);
2425     return RValue::get(Dest.getPointer());
2426   }
2427 
2428   case Builtin::BI__builtin_objc_memmove_collectable: {
2429     Address DestAddr = EmitPointerWithAlignment(E->getArg(0));
2430     Address SrcAddr = EmitPointerWithAlignment(E->getArg(1));
2431     Value *SizeVal = EmitScalarExpr(E->getArg(2));
2432     CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
2433                                                   DestAddr, SrcAddr, SizeVal);
2434     return RValue::get(DestAddr.getPointer());
2435   }
2436 
2437   case Builtin::BI__builtin___memmove_chk: {
2438     // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
2439     Expr::EvalResult SizeResult, DstSizeResult;
2440     if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
2441         !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
2442       break;
2443     llvm::APSInt Size = SizeResult.Val.getInt();
2444     llvm::APSInt DstSize = DstSizeResult.Val.getInt();
2445     if (Size.ugt(DstSize))
2446       break;
2447     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2448     Address Src = EmitPointerWithAlignment(E->getArg(1));
2449     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
2450     Builder.CreateMemMove(Dest, Src, SizeVal, false);
2451     return RValue::get(Dest.getPointer());
2452   }
2453 
2454   case Builtin::BImemmove:
2455   case Builtin::BI__builtin_memmove: {
2456     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2457     Address Src = EmitPointerWithAlignment(E->getArg(1));
2458     Value *SizeVal = EmitScalarExpr(E->getArg(2));
2459     EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2460                         E->getArg(0)->getExprLoc(), FD, 0);
2461     EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
2462                         E->getArg(1)->getExprLoc(), FD, 1);
2463     Builder.CreateMemMove(Dest, Src, SizeVal, false);
2464     return RValue::get(Dest.getPointer());
2465   }
2466   case Builtin::BImemset:
2467   case Builtin::BI__builtin_memset: {
2468     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2469     Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
2470                                          Builder.getInt8Ty());
2471     Value *SizeVal = EmitScalarExpr(E->getArg(2));
2472     EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2473                         E->getArg(0)->getExprLoc(), FD, 0);
2474     Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
2475     return RValue::get(Dest.getPointer());
2476   }
2477   case Builtin::BI__builtin___memset_chk: {
2478     // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
2479     Expr::EvalResult SizeResult, DstSizeResult;
2480     if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
2481         !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
2482       break;
2483     llvm::APSInt Size = SizeResult.Val.getInt();
2484     llvm::APSInt DstSize = DstSizeResult.Val.getInt();
2485     if (Size.ugt(DstSize))
2486       break;
2487     Address Dest = EmitPointerWithAlignment(E->getArg(0));
2488     Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
2489                                          Builder.getInt8Ty());
2490     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
2491     Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
2492     return RValue::get(Dest.getPointer());
2493   }
2494   case Builtin::BI__builtin_wmemcmp: {
2495     // The MSVC runtime library does not provide a definition of wmemcmp, so we
2496     // need an inline implementation.
2497     if (!getTarget().getTriple().isOSMSVCRT())
2498       break;
2499 
2500     llvm::Type *WCharTy = ConvertType(getContext().WCharTy);
2501 
2502     Value *Dst = EmitScalarExpr(E->getArg(0));
2503     Value *Src = EmitScalarExpr(E->getArg(1));
2504     Value *Size = EmitScalarExpr(E->getArg(2));
2505 
2506     BasicBlock *Entry = Builder.GetInsertBlock();
2507     BasicBlock *CmpGT = createBasicBlock("wmemcmp.gt");
2508     BasicBlock *CmpLT = createBasicBlock("wmemcmp.lt");
2509     BasicBlock *Next = createBasicBlock("wmemcmp.next");
2510     BasicBlock *Exit = createBasicBlock("wmemcmp.exit");
2511     Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0));
2512     Builder.CreateCondBr(SizeEq0, Exit, CmpGT);
2513 
2514     EmitBlock(CmpGT);
2515     PHINode *DstPhi = Builder.CreatePHI(Dst->getType(), 2);
2516     DstPhi->addIncoming(Dst, Entry);
2517     PHINode *SrcPhi = Builder.CreatePHI(Src->getType(), 2);
2518     SrcPhi->addIncoming(Src, Entry);
2519     PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2);
2520     SizePhi->addIncoming(Size, Entry);
2521     CharUnits WCharAlign =
2522         getContext().getTypeAlignInChars(getContext().WCharTy);
2523     Value *DstCh = Builder.CreateAlignedLoad(WCharTy, DstPhi, WCharAlign);
2524     Value *SrcCh = Builder.CreateAlignedLoad(WCharTy, SrcPhi, WCharAlign);
2525     Value *DstGtSrc = Builder.CreateICmpUGT(DstCh, SrcCh);
2526     Builder.CreateCondBr(DstGtSrc, Exit, CmpLT);
2527 
2528     EmitBlock(CmpLT);
2529     Value *DstLtSrc = Builder.CreateICmpULT(DstCh, SrcCh);
2530     Builder.CreateCondBr(DstLtSrc, Exit, Next);
2531 
2532     EmitBlock(Next);
2533     Value *NextDst = Builder.CreateConstInBoundsGEP1_32(WCharTy, DstPhi, 1);
2534     Value *NextSrc = Builder.CreateConstInBoundsGEP1_32(WCharTy, SrcPhi, 1);
2535     Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1));
2536     Value *NextSizeEq0 =
2537         Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0));
2538     Builder.CreateCondBr(NextSizeEq0, Exit, CmpGT);
2539     DstPhi->addIncoming(NextDst, Next);
2540     SrcPhi->addIncoming(NextSrc, Next);
2541     SizePhi->addIncoming(NextSize, Next);
2542 
2543     EmitBlock(Exit);
2544     PHINode *Ret = Builder.CreatePHI(IntTy, 4);
2545     Ret->addIncoming(ConstantInt::get(IntTy, 0), Entry);
2546     Ret->addIncoming(ConstantInt::get(IntTy, 1), CmpGT);
2547     Ret->addIncoming(ConstantInt::get(IntTy, -1), CmpLT);
2548     Ret->addIncoming(ConstantInt::get(IntTy, 0), Next);
2549     return RValue::get(Ret);
2550   }
2551   case Builtin::BI__builtin_dwarf_cfa: {
2552     // The offset in bytes from the first argument to the CFA.
2553     //
2554     // Why on earth is this in the frontend?  Is there any reason at
2555     // all that the backend can't reasonably determine this while
2556     // lowering llvm.eh.dwarf.cfa()?
2557     //
2558     // TODO: If there's a satisfactory reason, add a target hook for
2559     // this instead of hard-coding 0, which is correct for most targets.
2560     int32_t Offset = 0;
2561 
2562     Function *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
2563     return RValue::get(Builder.CreateCall(F,
2564                                       llvm::ConstantInt::get(Int32Ty, Offset)));
2565   }
2566   case Builtin::BI__builtin_return_address: {
2567     Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
2568                                                    getContext().UnsignedIntTy);
2569     Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
2570     return RValue::get(Builder.CreateCall(F, Depth));
2571   }
2572   case Builtin::BI_ReturnAddress: {
2573     Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
2574     return RValue::get(Builder.CreateCall(F, Builder.getInt32(0)));
2575   }
2576   case Builtin::BI__builtin_frame_address: {
2577     Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
2578                                                    getContext().UnsignedIntTy);
2579     Function *F = CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy);
2580     return RValue::get(Builder.CreateCall(F, Depth));
2581   }
2582   case Builtin::BI__builtin_extract_return_addr: {
2583     Value *Address = EmitScalarExpr(E->getArg(0));
2584     Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
2585     return RValue::get(Result);
2586   }
2587   case Builtin::BI__builtin_frob_return_addr: {
2588     Value *Address = EmitScalarExpr(E->getArg(0));
2589     Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
2590     return RValue::get(Result);
2591   }
2592   case Builtin::BI__builtin_dwarf_sp_column: {
2593     llvm::IntegerType *Ty
2594       = cast<llvm::IntegerType>(ConvertType(E->getType()));
2595     int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
2596     if (Column == -1) {
2597       CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
2598       return RValue::get(llvm::UndefValue::get(Ty));
2599     }
2600     return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
2601   }
2602   case Builtin::BI__builtin_init_dwarf_reg_size_table: {
2603     Value *Address = EmitScalarExpr(E->getArg(0));
2604     if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
2605       CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
2606     return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
2607   }
2608   case Builtin::BI__builtin_eh_return: {
2609     Value *Int = EmitScalarExpr(E->getArg(0));
2610     Value *Ptr = EmitScalarExpr(E->getArg(1));
2611 
2612     llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
2613     assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
2614            "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
2615     Function *F =
2616         CGM.getIntrinsic(IntTy->getBitWidth() == 32 ? Intrinsic::eh_return_i32
2617                                                     : Intrinsic::eh_return_i64);
2618     Builder.CreateCall(F, {Int, Ptr});
2619     Builder.CreateUnreachable();
2620 
2621     // We do need to preserve an insertion point.
2622     EmitBlock(createBasicBlock("builtin_eh_return.cont"));
2623 
2624     return RValue::get(nullptr);
2625   }
2626   case Builtin::BI__builtin_unwind_init: {
2627     Function *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
2628     return RValue::get(Builder.CreateCall(F));
2629   }
2630   case Builtin::BI__builtin_extend_pointer: {
2631     // Extends a pointer to the size of an _Unwind_Word, which is
2632     // uint64_t on all platforms.  Generally this gets poked into a
2633     // register and eventually used as an address, so if the
2634     // addressing registers are wider than pointers and the platform
2635     // doesn't implicitly ignore high-order bits when doing
2636     // addressing, we need to make sure we zext / sext based on
2637     // the platform's expectations.
2638     //
2639     // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
2640 
2641     // Cast the pointer to intptr_t.
2642     Value *Ptr = EmitScalarExpr(E->getArg(0));
2643     Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
2644 
2645     // If that's 64 bits, we're done.
2646     if (IntPtrTy->getBitWidth() == 64)
2647       return RValue::get(Result);
2648 
2649     // Otherwise, ask the codegen data what to do.
2650     if (getTargetHooks().extendPointerWithSExt())
2651       return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
2652     else
2653       return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
2654   }
2655   case Builtin::BI__builtin_setjmp: {
2656     // Buffer is a void**.
2657     Address Buf = EmitPointerWithAlignment(E->getArg(0));
2658 
2659     // Store the frame pointer to the setjmp buffer.
2660     Value *FrameAddr = Builder.CreateCall(
2661         CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy),
2662         ConstantInt::get(Int32Ty, 0));
2663     Builder.CreateStore(FrameAddr, Buf);
2664 
2665     // Store the stack pointer to the setjmp buffer.
2666     Value *StackAddr =
2667         Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
2668     Address StackSaveSlot = Builder.CreateConstInBoundsGEP(Buf, 2);
2669     Builder.CreateStore(StackAddr, StackSaveSlot);
2670 
2671     // Call LLVM's EH setjmp, which is lightweight.
2672     Function *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
2673     Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
2674     return RValue::get(Builder.CreateCall(F, Buf.getPointer()));
2675   }
2676   case Builtin::BI__builtin_longjmp: {
2677     Value *Buf = EmitScalarExpr(E->getArg(0));
2678     Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
2679 
2680     // Call LLVM's EH longjmp, which is lightweight.
2681     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
2682 
2683     // longjmp doesn't return; mark this as unreachable.
2684     Builder.CreateUnreachable();
2685 
2686     // We do need to preserve an insertion point.
2687     EmitBlock(createBasicBlock("longjmp.cont"));
2688 
2689     return RValue::get(nullptr);
2690   }
2691   case Builtin::BI__builtin_launder: {
2692     const Expr *Arg = E->getArg(0);
2693     QualType ArgTy = Arg->getType()->getPointeeType();
2694     Value *Ptr = EmitScalarExpr(Arg);
2695     if (TypeRequiresBuiltinLaunder(CGM, ArgTy))
2696       Ptr = Builder.CreateLaunderInvariantGroup(Ptr);
2697 
2698     return RValue::get(Ptr);
2699   }
2700   case Builtin::BI__sync_fetch_and_add:
2701   case Builtin::BI__sync_fetch_and_sub:
2702   case Builtin::BI__sync_fetch_and_or:
2703   case Builtin::BI__sync_fetch_and_and:
2704   case Builtin::BI__sync_fetch_and_xor:
2705   case Builtin::BI__sync_fetch_and_nand:
2706   case Builtin::BI__sync_add_and_fetch:
2707   case Builtin::BI__sync_sub_and_fetch:
2708   case Builtin::BI__sync_and_and_fetch:
2709   case Builtin::BI__sync_or_and_fetch:
2710   case Builtin::BI__sync_xor_and_fetch:
2711   case Builtin::BI__sync_nand_and_fetch:
2712   case Builtin::BI__sync_val_compare_and_swap:
2713   case Builtin::BI__sync_bool_compare_and_swap:
2714   case Builtin::BI__sync_lock_test_and_set:
2715   case Builtin::BI__sync_lock_release:
2716   case Builtin::BI__sync_swap:
2717     llvm_unreachable("Shouldn't make it through sema");
2718   case Builtin::BI__sync_fetch_and_add_1:
2719   case Builtin::BI__sync_fetch_and_add_2:
2720   case Builtin::BI__sync_fetch_and_add_4:
2721   case Builtin::BI__sync_fetch_and_add_8:
2722   case Builtin::BI__sync_fetch_and_add_16:
2723     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
2724   case Builtin::BI__sync_fetch_and_sub_1:
2725   case Builtin::BI__sync_fetch_and_sub_2:
2726   case Builtin::BI__sync_fetch_and_sub_4:
2727   case Builtin::BI__sync_fetch_and_sub_8:
2728   case Builtin::BI__sync_fetch_and_sub_16:
2729     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
2730   case Builtin::BI__sync_fetch_and_or_1:
2731   case Builtin::BI__sync_fetch_and_or_2:
2732   case Builtin::BI__sync_fetch_and_or_4:
2733   case Builtin::BI__sync_fetch_and_or_8:
2734   case Builtin::BI__sync_fetch_and_or_16:
2735     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
2736   case Builtin::BI__sync_fetch_and_and_1:
2737   case Builtin::BI__sync_fetch_and_and_2:
2738   case Builtin::BI__sync_fetch_and_and_4:
2739   case Builtin::BI__sync_fetch_and_and_8:
2740   case Builtin::BI__sync_fetch_and_and_16:
2741     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
2742   case Builtin::BI__sync_fetch_and_xor_1:
2743   case Builtin::BI__sync_fetch_and_xor_2:
2744   case Builtin::BI__sync_fetch_and_xor_4:
2745   case Builtin::BI__sync_fetch_and_xor_8:
2746   case Builtin::BI__sync_fetch_and_xor_16:
2747     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
2748   case Builtin::BI__sync_fetch_and_nand_1:
2749   case Builtin::BI__sync_fetch_and_nand_2:
2750   case Builtin::BI__sync_fetch_and_nand_4:
2751   case Builtin::BI__sync_fetch_and_nand_8:
2752   case Builtin::BI__sync_fetch_and_nand_16:
2753     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E);
2754 
2755   // Clang extensions: not overloaded yet.
2756   case Builtin::BI__sync_fetch_and_min:
2757     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
2758   case Builtin::BI__sync_fetch_and_max:
2759     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
2760   case Builtin::BI__sync_fetch_and_umin:
2761     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
2762   case Builtin::BI__sync_fetch_and_umax:
2763     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
2764 
2765   case Builtin::BI__sync_add_and_fetch_1:
2766   case Builtin::BI__sync_add_and_fetch_2:
2767   case Builtin::BI__sync_add_and_fetch_4:
2768   case Builtin::BI__sync_add_and_fetch_8:
2769   case Builtin::BI__sync_add_and_fetch_16:
2770     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
2771                                 llvm::Instruction::Add);
2772   case Builtin::BI__sync_sub_and_fetch_1:
2773   case Builtin::BI__sync_sub_and_fetch_2:
2774   case Builtin::BI__sync_sub_and_fetch_4:
2775   case Builtin::BI__sync_sub_and_fetch_8:
2776   case Builtin::BI__sync_sub_and_fetch_16:
2777     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
2778                                 llvm::Instruction::Sub);
2779   case Builtin::BI__sync_and_and_fetch_1:
2780   case Builtin::BI__sync_and_and_fetch_2:
2781   case Builtin::BI__sync_and_and_fetch_4:
2782   case Builtin::BI__sync_and_and_fetch_8:
2783   case Builtin::BI__sync_and_and_fetch_16:
2784     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
2785                                 llvm::Instruction::And);
2786   case Builtin::BI__sync_or_and_fetch_1:
2787   case Builtin::BI__sync_or_and_fetch_2:
2788   case Builtin::BI__sync_or_and_fetch_4:
2789   case Builtin::BI__sync_or_and_fetch_8:
2790   case Builtin::BI__sync_or_and_fetch_16:
2791     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
2792                                 llvm::Instruction::Or);
2793   case Builtin::BI__sync_xor_and_fetch_1:
2794   case Builtin::BI__sync_xor_and_fetch_2:
2795   case Builtin::BI__sync_xor_and_fetch_4:
2796   case Builtin::BI__sync_xor_and_fetch_8:
2797   case Builtin::BI__sync_xor_and_fetch_16:
2798     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
2799                                 llvm::Instruction::Xor);
2800   case Builtin::BI__sync_nand_and_fetch_1:
2801   case Builtin::BI__sync_nand_and_fetch_2:
2802   case Builtin::BI__sync_nand_and_fetch_4:
2803   case Builtin::BI__sync_nand_and_fetch_8:
2804   case Builtin::BI__sync_nand_and_fetch_16:
2805     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E,
2806                                 llvm::Instruction::And, true);
2807 
2808   case Builtin::BI__sync_val_compare_and_swap_1:
2809   case Builtin::BI__sync_val_compare_and_swap_2:
2810   case Builtin::BI__sync_val_compare_and_swap_4:
2811   case Builtin::BI__sync_val_compare_and_swap_8:
2812   case Builtin::BI__sync_val_compare_and_swap_16:
2813     return RValue::get(MakeAtomicCmpXchgValue(*this, E, false));
2814 
2815   case Builtin::BI__sync_bool_compare_and_swap_1:
2816   case Builtin::BI__sync_bool_compare_and_swap_2:
2817   case Builtin::BI__sync_bool_compare_and_swap_4:
2818   case Builtin::BI__sync_bool_compare_and_swap_8:
2819   case Builtin::BI__sync_bool_compare_and_swap_16:
2820     return RValue::get(MakeAtomicCmpXchgValue(*this, E, true));
2821 
2822   case Builtin::BI__sync_swap_1:
2823   case Builtin::BI__sync_swap_2:
2824   case Builtin::BI__sync_swap_4:
2825   case Builtin::BI__sync_swap_8:
2826   case Builtin::BI__sync_swap_16:
2827     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
2828 
2829   case Builtin::BI__sync_lock_test_and_set_1:
2830   case Builtin::BI__sync_lock_test_and_set_2:
2831   case Builtin::BI__sync_lock_test_and_set_4:
2832   case Builtin::BI__sync_lock_test_and_set_8:
2833   case Builtin::BI__sync_lock_test_and_set_16:
2834     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
2835 
2836   case Builtin::BI__sync_lock_release_1:
2837   case Builtin::BI__sync_lock_release_2:
2838   case Builtin::BI__sync_lock_release_4:
2839   case Builtin::BI__sync_lock_release_8:
2840   case Builtin::BI__sync_lock_release_16: {
2841     Value *Ptr = EmitScalarExpr(E->getArg(0));
2842     QualType ElTy = E->getArg(0)->getType()->getPointeeType();
2843     CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
2844     llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
2845                                              StoreSize.getQuantity() * 8);
2846     Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
2847     llvm::StoreInst *Store =
2848       Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr,
2849                                  StoreSize);
2850     Store->setAtomic(llvm::AtomicOrdering::Release);
2851     return RValue::get(nullptr);
2852   }
2853 
2854   case Builtin::BI__sync_synchronize: {
2855     // We assume this is supposed to correspond to a C++0x-style
2856     // sequentially-consistent fence (i.e. this is only usable for
2857     // synchronization, not device I/O or anything like that). This intrinsic
2858     // is really badly designed in the sense that in theory, there isn't
2859     // any way to safely use it... but in practice, it mostly works
2860     // to use it with non-atomic loads and stores to get acquire/release
2861     // semantics.
2862     Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent);
2863     return RValue::get(nullptr);
2864   }
2865 
2866   case Builtin::BI__builtin_nontemporal_load:
2867     return RValue::get(EmitNontemporalLoad(*this, E));
2868   case Builtin::BI__builtin_nontemporal_store:
2869     return RValue::get(EmitNontemporalStore(*this, E));
2870   case Builtin::BI__c11_atomic_is_lock_free:
2871   case Builtin::BI__atomic_is_lock_free: {
2872     // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
2873     // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
2874     // _Atomic(T) is always properly-aligned.
2875     const char *LibCallName = "__atomic_is_lock_free";
2876     CallArgList Args;
2877     Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
2878              getContext().getSizeType());
2879     if (BuiltinID == Builtin::BI__atomic_is_lock_free)
2880       Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
2881                getContext().VoidPtrTy);
2882     else
2883       Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
2884                getContext().VoidPtrTy);
2885     const CGFunctionInfo &FuncInfo =
2886         CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args);
2887     llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
2888     llvm::FunctionCallee Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
2889     return EmitCall(FuncInfo, CGCallee::forDirect(Func),
2890                     ReturnValueSlot(), Args);
2891   }
2892 
2893   case Builtin::BI__atomic_test_and_set: {
2894     // Look at the argument type to determine whether this is a volatile
2895     // operation. The parameter type is always volatile.
2896     QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
2897     bool Volatile =
2898         PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
2899 
2900     Value *Ptr = EmitScalarExpr(E->getArg(0));
2901     unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
2902     Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
2903     Value *NewVal = Builder.getInt8(1);
2904     Value *Order = EmitScalarExpr(E->getArg(1));
2905     if (isa<llvm::ConstantInt>(Order)) {
2906       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
2907       AtomicRMWInst *Result = nullptr;
2908       switch (ord) {
2909       case 0:  // memory_order_relaxed
2910       default: // invalid order
2911         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2912                                          llvm::AtomicOrdering::Monotonic);
2913         break;
2914       case 1: // memory_order_consume
2915       case 2: // memory_order_acquire
2916         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2917                                          llvm::AtomicOrdering::Acquire);
2918         break;
2919       case 3: // memory_order_release
2920         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2921                                          llvm::AtomicOrdering::Release);
2922         break;
2923       case 4: // memory_order_acq_rel
2924 
2925         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2926                                          llvm::AtomicOrdering::AcquireRelease);
2927         break;
2928       case 5: // memory_order_seq_cst
2929         Result = Builder.CreateAtomicRMW(
2930             llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2931             llvm::AtomicOrdering::SequentiallyConsistent);
2932         break;
2933       }
2934       Result->setVolatile(Volatile);
2935       return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
2936     }
2937 
2938     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
2939 
2940     llvm::BasicBlock *BBs[5] = {
2941       createBasicBlock("monotonic", CurFn),
2942       createBasicBlock("acquire", CurFn),
2943       createBasicBlock("release", CurFn),
2944       createBasicBlock("acqrel", CurFn),
2945       createBasicBlock("seqcst", CurFn)
2946     };
2947     llvm::AtomicOrdering Orders[5] = {
2948         llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire,
2949         llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease,
2950         llvm::AtomicOrdering::SequentiallyConsistent};
2951 
2952     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
2953     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
2954 
2955     Builder.SetInsertPoint(ContBB);
2956     PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
2957 
2958     for (unsigned i = 0; i < 5; ++i) {
2959       Builder.SetInsertPoint(BBs[i]);
2960       AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
2961                                                    Ptr, NewVal, Orders[i]);
2962       RMW->setVolatile(Volatile);
2963       Result->addIncoming(RMW, BBs[i]);
2964       Builder.CreateBr(ContBB);
2965     }
2966 
2967     SI->addCase(Builder.getInt32(0), BBs[0]);
2968     SI->addCase(Builder.getInt32(1), BBs[1]);
2969     SI->addCase(Builder.getInt32(2), BBs[1]);
2970     SI->addCase(Builder.getInt32(3), BBs[2]);
2971     SI->addCase(Builder.getInt32(4), BBs[3]);
2972     SI->addCase(Builder.getInt32(5), BBs[4]);
2973 
2974     Builder.SetInsertPoint(ContBB);
2975     return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
2976   }
2977 
2978   case Builtin::BI__atomic_clear: {
2979     QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
2980     bool Volatile =
2981         PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
2982 
2983     Address Ptr = EmitPointerWithAlignment(E->getArg(0));
2984     unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace();
2985     Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
2986     Value *NewVal = Builder.getInt8(0);
2987     Value *Order = EmitScalarExpr(E->getArg(1));
2988     if (isa<llvm::ConstantInt>(Order)) {
2989       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
2990       StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
2991       switch (ord) {
2992       case 0:  // memory_order_relaxed
2993       default: // invalid order
2994         Store->setOrdering(llvm::AtomicOrdering::Monotonic);
2995         break;
2996       case 3:  // memory_order_release
2997         Store->setOrdering(llvm::AtomicOrdering::Release);
2998         break;
2999       case 5:  // memory_order_seq_cst
3000         Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent);
3001         break;
3002       }
3003       return RValue::get(nullptr);
3004     }
3005 
3006     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
3007 
3008     llvm::BasicBlock *BBs[3] = {
3009       createBasicBlock("monotonic", CurFn),
3010       createBasicBlock("release", CurFn),
3011       createBasicBlock("seqcst", CurFn)
3012     };
3013     llvm::AtomicOrdering Orders[3] = {
3014         llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release,
3015         llvm::AtomicOrdering::SequentiallyConsistent};
3016 
3017     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
3018     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
3019 
3020     for (unsigned i = 0; i < 3; ++i) {
3021       Builder.SetInsertPoint(BBs[i]);
3022       StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
3023       Store->setOrdering(Orders[i]);
3024       Builder.CreateBr(ContBB);
3025     }
3026 
3027     SI->addCase(Builder.getInt32(0), BBs[0]);
3028     SI->addCase(Builder.getInt32(3), BBs[1]);
3029     SI->addCase(Builder.getInt32(5), BBs[2]);
3030 
3031     Builder.SetInsertPoint(ContBB);
3032     return RValue::get(nullptr);
3033   }
3034 
3035   case Builtin::BI__atomic_thread_fence:
3036   case Builtin::BI__atomic_signal_fence:
3037   case Builtin::BI__c11_atomic_thread_fence:
3038   case Builtin::BI__c11_atomic_signal_fence: {
3039     llvm::SyncScope::ID SSID;
3040     if (BuiltinID == Builtin::BI__atomic_signal_fence ||
3041         BuiltinID == Builtin::BI__c11_atomic_signal_fence)
3042       SSID = llvm::SyncScope::SingleThread;
3043     else
3044       SSID = llvm::SyncScope::System;
3045     Value *Order = EmitScalarExpr(E->getArg(0));
3046     if (isa<llvm::ConstantInt>(Order)) {
3047       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
3048       switch (ord) {
3049       case 0:  // memory_order_relaxed
3050       default: // invalid order
3051         break;
3052       case 1:  // memory_order_consume
3053       case 2:  // memory_order_acquire
3054         Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
3055         break;
3056       case 3:  // memory_order_release
3057         Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
3058         break;
3059       case 4:  // memory_order_acq_rel
3060         Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
3061         break;
3062       case 5:  // memory_order_seq_cst
3063         Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
3064         break;
3065       }
3066       return RValue::get(nullptr);
3067     }
3068 
3069     llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
3070     AcquireBB = createBasicBlock("acquire", CurFn);
3071     ReleaseBB = createBasicBlock("release", CurFn);
3072     AcqRelBB = createBasicBlock("acqrel", CurFn);
3073     SeqCstBB = createBasicBlock("seqcst", CurFn);
3074     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
3075 
3076     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
3077     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
3078 
3079     Builder.SetInsertPoint(AcquireBB);
3080     Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
3081     Builder.CreateBr(ContBB);
3082     SI->addCase(Builder.getInt32(1), AcquireBB);
3083     SI->addCase(Builder.getInt32(2), AcquireBB);
3084 
3085     Builder.SetInsertPoint(ReleaseBB);
3086     Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
3087     Builder.CreateBr(ContBB);
3088     SI->addCase(Builder.getInt32(3), ReleaseBB);
3089 
3090     Builder.SetInsertPoint(AcqRelBB);
3091     Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
3092     Builder.CreateBr(ContBB);
3093     SI->addCase(Builder.getInt32(4), AcqRelBB);
3094 
3095     Builder.SetInsertPoint(SeqCstBB);
3096     Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
3097     Builder.CreateBr(ContBB);
3098     SI->addCase(Builder.getInt32(5), SeqCstBB);
3099 
3100     Builder.SetInsertPoint(ContBB);
3101     return RValue::get(nullptr);
3102   }
3103 
3104   case Builtin::BI__builtin_signbit:
3105   case Builtin::BI__builtin_signbitf:
3106   case Builtin::BI__builtin_signbitl: {
3107     return RValue::get(
3108         Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))),
3109                            ConvertType(E->getType())));
3110   }
3111   case Builtin::BI__annotation: {
3112     // Re-encode each wide string to UTF8 and make an MDString.
3113     SmallVector<Metadata *, 1> Strings;
3114     for (const Expr *Arg : E->arguments()) {
3115       const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts());
3116       assert(Str->getCharByteWidth() == 2);
3117       StringRef WideBytes = Str->getBytes();
3118       std::string StrUtf8;
3119       if (!convertUTF16ToUTF8String(
3120               makeArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) {
3121         CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument");
3122         continue;
3123       }
3124       Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8));
3125     }
3126 
3127     // Build and MDTuple of MDStrings and emit the intrinsic call.
3128     llvm::Function *F =
3129         CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {});
3130     MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings);
3131     Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple));
3132     return RValue::getIgnored();
3133   }
3134   case Builtin::BI__builtin_annotation: {
3135     llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
3136     llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
3137                                       AnnVal->getType());
3138 
3139     // Get the annotation string, go through casts. Sema requires this to be a
3140     // non-wide string literal, potentially casted, so the cast<> is safe.
3141     const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
3142     StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
3143     return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
3144   }
3145   case Builtin::BI__builtin_addcb:
3146   case Builtin::BI__builtin_addcs:
3147   case Builtin::BI__builtin_addc:
3148   case Builtin::BI__builtin_addcl:
3149   case Builtin::BI__builtin_addcll:
3150   case Builtin::BI__builtin_subcb:
3151   case Builtin::BI__builtin_subcs:
3152   case Builtin::BI__builtin_subc:
3153   case Builtin::BI__builtin_subcl:
3154   case Builtin::BI__builtin_subcll: {
3155 
3156     // We translate all of these builtins from expressions of the form:
3157     //   int x = ..., y = ..., carryin = ..., carryout, result;
3158     //   result = __builtin_addc(x, y, carryin, &carryout);
3159     //
3160     // to LLVM IR of the form:
3161     //
3162     //   %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
3163     //   %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
3164     //   %carry1 = extractvalue {i32, i1} %tmp1, 1
3165     //   %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
3166     //                                                       i32 %carryin)
3167     //   %result = extractvalue {i32, i1} %tmp2, 0
3168     //   %carry2 = extractvalue {i32, i1} %tmp2, 1
3169     //   %tmp3 = or i1 %carry1, %carry2
3170     //   %tmp4 = zext i1 %tmp3 to i32
3171     //   store i32 %tmp4, i32* %carryout
3172 
3173     // Scalarize our inputs.
3174     llvm::Value *X = EmitScalarExpr(E->getArg(0));
3175     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
3176     llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
3177     Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3));
3178 
3179     // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
3180     llvm::Intrinsic::ID IntrinsicId;
3181     switch (BuiltinID) {
3182     default: llvm_unreachable("Unknown multiprecision builtin id.");
3183     case Builtin::BI__builtin_addcb:
3184     case Builtin::BI__builtin_addcs:
3185     case Builtin::BI__builtin_addc:
3186     case Builtin::BI__builtin_addcl:
3187     case Builtin::BI__builtin_addcll:
3188       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
3189       break;
3190     case Builtin::BI__builtin_subcb:
3191     case Builtin::BI__builtin_subcs:
3192     case Builtin::BI__builtin_subc:
3193     case Builtin::BI__builtin_subcl:
3194     case Builtin::BI__builtin_subcll:
3195       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
3196       break;
3197     }
3198 
3199     // Construct our resulting LLVM IR expression.
3200     llvm::Value *Carry1;
3201     llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
3202                                               X, Y, Carry1);
3203     llvm::Value *Carry2;
3204     llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
3205                                               Sum1, Carryin, Carry2);
3206     llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
3207                                                X->getType());
3208     Builder.CreateStore(CarryOut, CarryOutPtr);
3209     return RValue::get(Sum2);
3210   }
3211 
3212   case Builtin::BI__builtin_add_overflow:
3213   case Builtin::BI__builtin_sub_overflow:
3214   case Builtin::BI__builtin_mul_overflow: {
3215     const clang::Expr *LeftArg = E->getArg(0);
3216     const clang::Expr *RightArg = E->getArg(1);
3217     const clang::Expr *ResultArg = E->getArg(2);
3218 
3219     clang::QualType ResultQTy =
3220         ResultArg->getType()->castAs<PointerType>()->getPointeeType();
3221 
3222     WidthAndSignedness LeftInfo =
3223         getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType());
3224     WidthAndSignedness RightInfo =
3225         getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType());
3226     WidthAndSignedness ResultInfo =
3227         getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy);
3228 
3229     // Handle mixed-sign multiplication as a special case, because adding
3230     // runtime or backend support for our generic irgen would be too expensive.
3231     if (isSpecialMixedSignMultiply(BuiltinID, LeftInfo, RightInfo, ResultInfo))
3232       return EmitCheckedMixedSignMultiply(*this, LeftArg, LeftInfo, RightArg,
3233                                           RightInfo, ResultArg, ResultQTy,
3234                                           ResultInfo);
3235 
3236     WidthAndSignedness EncompassingInfo =
3237         EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo});
3238 
3239     llvm::Type *EncompassingLLVMTy =
3240         llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width);
3241 
3242     llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy);
3243 
3244     llvm::Intrinsic::ID IntrinsicId;
3245     switch (BuiltinID) {
3246     default:
3247       llvm_unreachable("Unknown overflow builtin id.");
3248     case Builtin::BI__builtin_add_overflow:
3249       IntrinsicId = EncompassingInfo.Signed
3250                         ? llvm::Intrinsic::sadd_with_overflow
3251                         : llvm::Intrinsic::uadd_with_overflow;
3252       break;
3253     case Builtin::BI__builtin_sub_overflow:
3254       IntrinsicId = EncompassingInfo.Signed
3255                         ? llvm::Intrinsic::ssub_with_overflow
3256                         : llvm::Intrinsic::usub_with_overflow;
3257       break;
3258     case Builtin::BI__builtin_mul_overflow:
3259       IntrinsicId = EncompassingInfo.Signed
3260                         ? llvm::Intrinsic::smul_with_overflow
3261                         : llvm::Intrinsic::umul_with_overflow;
3262       break;
3263     }
3264 
3265     llvm::Value *Left = EmitScalarExpr(LeftArg);
3266     llvm::Value *Right = EmitScalarExpr(RightArg);
3267     Address ResultPtr = EmitPointerWithAlignment(ResultArg);
3268 
3269     // Extend each operand to the encompassing type.
3270     Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed);
3271     Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed);
3272 
3273     // Perform the operation on the extended values.
3274     llvm::Value *Overflow, *Result;
3275     Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow);
3276 
3277     if (EncompassingInfo.Width > ResultInfo.Width) {
3278       // The encompassing type is wider than the result type, so we need to
3279       // truncate it.
3280       llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy);
3281 
3282       // To see if the truncation caused an overflow, we will extend
3283       // the result and then compare it to the original result.
3284       llvm::Value *ResultTruncExt = Builder.CreateIntCast(
3285           ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed);
3286       llvm::Value *TruncationOverflow =
3287           Builder.CreateICmpNE(Result, ResultTruncExt);
3288 
3289       Overflow = Builder.CreateOr(Overflow, TruncationOverflow);
3290       Result = ResultTrunc;
3291     }
3292 
3293     // Finally, store the result using the pointer.
3294     bool isVolatile =
3295       ResultArg->getType()->getPointeeType().isVolatileQualified();
3296     Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile);
3297 
3298     return RValue::get(Overflow);
3299   }
3300 
3301   case Builtin::BI__builtin_uadd_overflow:
3302   case Builtin::BI__builtin_uaddl_overflow:
3303   case Builtin::BI__builtin_uaddll_overflow:
3304   case Builtin::BI__builtin_usub_overflow:
3305   case Builtin::BI__builtin_usubl_overflow:
3306   case Builtin::BI__builtin_usubll_overflow:
3307   case Builtin::BI__builtin_umul_overflow:
3308   case Builtin::BI__builtin_umull_overflow:
3309   case Builtin::BI__builtin_umulll_overflow:
3310   case Builtin::BI__builtin_sadd_overflow:
3311   case Builtin::BI__builtin_saddl_overflow:
3312   case Builtin::BI__builtin_saddll_overflow:
3313   case Builtin::BI__builtin_ssub_overflow:
3314   case Builtin::BI__builtin_ssubl_overflow:
3315   case Builtin::BI__builtin_ssubll_overflow:
3316   case Builtin::BI__builtin_smul_overflow:
3317   case Builtin::BI__builtin_smull_overflow:
3318   case Builtin::BI__builtin_smulll_overflow: {
3319 
3320     // We translate all of these builtins directly to the relevant llvm IR node.
3321 
3322     // Scalarize our inputs.
3323     llvm::Value *X = EmitScalarExpr(E->getArg(0));
3324     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
3325     Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2));
3326 
3327     // Decide which of the overflow intrinsics we are lowering to:
3328     llvm::Intrinsic::ID IntrinsicId;
3329     switch (BuiltinID) {
3330     default: llvm_unreachable("Unknown overflow builtin id.");
3331     case Builtin::BI__builtin_uadd_overflow:
3332     case Builtin::BI__builtin_uaddl_overflow:
3333     case Builtin::BI__builtin_uaddll_overflow:
3334       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
3335       break;
3336     case Builtin::BI__builtin_usub_overflow:
3337     case Builtin::BI__builtin_usubl_overflow:
3338     case Builtin::BI__builtin_usubll_overflow:
3339       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
3340       break;
3341     case Builtin::BI__builtin_umul_overflow:
3342     case Builtin::BI__builtin_umull_overflow:
3343     case Builtin::BI__builtin_umulll_overflow:
3344       IntrinsicId = llvm::Intrinsic::umul_with_overflow;
3345       break;
3346     case Builtin::BI__builtin_sadd_overflow:
3347     case Builtin::BI__builtin_saddl_overflow:
3348     case Builtin::BI__builtin_saddll_overflow:
3349       IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
3350       break;
3351     case Builtin::BI__builtin_ssub_overflow:
3352     case Builtin::BI__builtin_ssubl_overflow:
3353     case Builtin::BI__builtin_ssubll_overflow:
3354       IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
3355       break;
3356     case Builtin::BI__builtin_smul_overflow:
3357     case Builtin::BI__builtin_smull_overflow:
3358     case Builtin::BI__builtin_smulll_overflow:
3359       IntrinsicId = llvm::Intrinsic::smul_with_overflow;
3360       break;
3361     }
3362 
3363 
3364     llvm::Value *Carry;
3365     llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
3366     Builder.CreateStore(Sum, SumOutPtr);
3367 
3368     return RValue::get(Carry);
3369   }
3370   case Builtin::BI__builtin_addressof:
3371     return RValue::get(EmitLValue(E->getArg(0)).getPointer(*this));
3372   case Builtin::BI__builtin_operator_new:
3373     return EmitBuiltinNewDeleteCall(
3374         E->getCallee()->getType()->castAs<FunctionProtoType>(), E, false);
3375   case Builtin::BI__builtin_operator_delete:
3376     return EmitBuiltinNewDeleteCall(
3377         E->getCallee()->getType()->castAs<FunctionProtoType>(), E, true);
3378 
3379   case Builtin::BI__noop:
3380     // __noop always evaluates to an integer literal zero.
3381     return RValue::get(ConstantInt::get(IntTy, 0));
3382   case Builtin::BI__builtin_call_with_static_chain: {
3383     const CallExpr *Call = cast<CallExpr>(E->getArg(0));
3384     const Expr *Chain = E->getArg(1);
3385     return EmitCall(Call->getCallee()->getType(),
3386                     EmitCallee(Call->getCallee()), Call, ReturnValue,
3387                     EmitScalarExpr(Chain));
3388   }
3389   case Builtin::BI_InterlockedExchange8:
3390   case Builtin::BI_InterlockedExchange16:
3391   case Builtin::BI_InterlockedExchange:
3392   case Builtin::BI_InterlockedExchangePointer:
3393     return RValue::get(
3394         EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E));
3395   case Builtin::BI_InterlockedCompareExchangePointer:
3396   case Builtin::BI_InterlockedCompareExchangePointer_nf: {
3397     llvm::Type *RTy;
3398     llvm::IntegerType *IntType =
3399       IntegerType::get(getLLVMContext(),
3400                        getContext().getTypeSize(E->getType()));
3401     llvm::Type *IntPtrType = IntType->getPointerTo();
3402 
3403     llvm::Value *Destination =
3404       Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType);
3405 
3406     llvm::Value *Exchange = EmitScalarExpr(E->getArg(1));
3407     RTy = Exchange->getType();
3408     Exchange = Builder.CreatePtrToInt(Exchange, IntType);
3409 
3410     llvm::Value *Comparand =
3411       Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType);
3412 
3413     auto Ordering =
3414       BuiltinID == Builtin::BI_InterlockedCompareExchangePointer_nf ?
3415       AtomicOrdering::Monotonic : AtomicOrdering::SequentiallyConsistent;
3416 
3417     auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
3418                                               Ordering, Ordering);
3419     Result->setVolatile(true);
3420 
3421     return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result,
3422                                                                          0),
3423                                               RTy));
3424   }
3425   case Builtin::BI_InterlockedCompareExchange8:
3426   case Builtin::BI_InterlockedCompareExchange16:
3427   case Builtin::BI_InterlockedCompareExchange:
3428   case Builtin::BI_InterlockedCompareExchange64:
3429     return RValue::get(EmitAtomicCmpXchgForMSIntrin(*this, E));
3430   case Builtin::BI_InterlockedIncrement16:
3431   case Builtin::BI_InterlockedIncrement:
3432     return RValue::get(
3433         EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E));
3434   case Builtin::BI_InterlockedDecrement16:
3435   case Builtin::BI_InterlockedDecrement:
3436     return RValue::get(
3437         EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E));
3438   case Builtin::BI_InterlockedAnd8:
3439   case Builtin::BI_InterlockedAnd16:
3440   case Builtin::BI_InterlockedAnd:
3441     return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E));
3442   case Builtin::BI_InterlockedExchangeAdd8:
3443   case Builtin::BI_InterlockedExchangeAdd16:
3444   case Builtin::BI_InterlockedExchangeAdd:
3445     return RValue::get(
3446         EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E));
3447   case Builtin::BI_InterlockedExchangeSub8:
3448   case Builtin::BI_InterlockedExchangeSub16:
3449   case Builtin::BI_InterlockedExchangeSub:
3450     return RValue::get(
3451         EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E));
3452   case Builtin::BI_InterlockedOr8:
3453   case Builtin::BI_InterlockedOr16:
3454   case Builtin::BI_InterlockedOr:
3455     return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E));
3456   case Builtin::BI_InterlockedXor8:
3457   case Builtin::BI_InterlockedXor16:
3458   case Builtin::BI_InterlockedXor:
3459     return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E));
3460 
3461   case Builtin::BI_bittest64:
3462   case Builtin::BI_bittest:
3463   case Builtin::BI_bittestandcomplement64:
3464   case Builtin::BI_bittestandcomplement:
3465   case Builtin::BI_bittestandreset64:
3466   case Builtin::BI_bittestandreset:
3467   case Builtin::BI_bittestandset64:
3468   case Builtin::BI_bittestandset:
3469   case Builtin::BI_interlockedbittestandreset:
3470   case Builtin::BI_interlockedbittestandreset64:
3471   case Builtin::BI_interlockedbittestandset64:
3472   case Builtin::BI_interlockedbittestandset:
3473   case Builtin::BI_interlockedbittestandset_acq:
3474   case Builtin::BI_interlockedbittestandset_rel:
3475   case Builtin::BI_interlockedbittestandset_nf:
3476   case Builtin::BI_interlockedbittestandreset_acq:
3477   case Builtin::BI_interlockedbittestandreset_rel:
3478   case Builtin::BI_interlockedbittestandreset_nf:
3479     return RValue::get(EmitBitTestIntrinsic(*this, BuiltinID, E));
3480 
3481     // These builtins exist to emit regular volatile loads and stores not
3482     // affected by the -fms-volatile setting.
3483   case Builtin::BI__iso_volatile_load8:
3484   case Builtin::BI__iso_volatile_load16:
3485   case Builtin::BI__iso_volatile_load32:
3486   case Builtin::BI__iso_volatile_load64:
3487     return RValue::get(EmitISOVolatileLoad(*this, E));
3488   case Builtin::BI__iso_volatile_store8:
3489   case Builtin::BI__iso_volatile_store16:
3490   case Builtin::BI__iso_volatile_store32:
3491   case Builtin::BI__iso_volatile_store64:
3492     return RValue::get(EmitISOVolatileStore(*this, E));
3493 
3494   case Builtin::BI__exception_code:
3495   case Builtin::BI_exception_code:
3496     return RValue::get(EmitSEHExceptionCode());
3497   case Builtin::BI__exception_info:
3498   case Builtin::BI_exception_info:
3499     return RValue::get(EmitSEHExceptionInfo());
3500   case Builtin::BI__abnormal_termination:
3501   case Builtin::BI_abnormal_termination:
3502     return RValue::get(EmitSEHAbnormalTermination());
3503   case Builtin::BI_setjmpex:
3504     if (getTarget().getTriple().isOSMSVCRT())
3505       return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
3506     break;
3507   case Builtin::BI_setjmp:
3508     if (getTarget().getTriple().isOSMSVCRT()) {
3509       if (getTarget().getTriple().getArch() == llvm::Triple::x86)
3510         return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp3, E);
3511       else if (getTarget().getTriple().getArch() == llvm::Triple::aarch64)
3512         return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
3513       return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp, E);
3514     }
3515     break;
3516 
3517   case Builtin::BI__GetExceptionInfo: {
3518     if (llvm::GlobalVariable *GV =
3519             CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType()))
3520       return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy));
3521     break;
3522   }
3523 
3524   case Builtin::BI__fastfail:
3525     return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E));
3526 
3527   case Builtin::BI__builtin_coro_size: {
3528     auto & Context = getContext();
3529     auto SizeTy = Context.getSizeType();
3530     auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3531     Function *F = CGM.getIntrinsic(Intrinsic::coro_size, T);
3532     return RValue::get(Builder.CreateCall(F));
3533   }
3534 
3535   case Builtin::BI__builtin_coro_id:
3536     return EmitCoroutineIntrinsic(E, Intrinsic::coro_id);
3537   case Builtin::BI__builtin_coro_promise:
3538     return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise);
3539   case Builtin::BI__builtin_coro_resume:
3540     return EmitCoroutineIntrinsic(E, Intrinsic::coro_resume);
3541   case Builtin::BI__builtin_coro_frame:
3542     return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame);
3543   case Builtin::BI__builtin_coro_noop:
3544     return EmitCoroutineIntrinsic(E, Intrinsic::coro_noop);
3545   case Builtin::BI__builtin_coro_free:
3546     return EmitCoroutineIntrinsic(E, Intrinsic::coro_free);
3547   case Builtin::BI__builtin_coro_destroy:
3548     return EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy);
3549   case Builtin::BI__builtin_coro_done:
3550     return EmitCoroutineIntrinsic(E, Intrinsic::coro_done);
3551   case Builtin::BI__builtin_coro_alloc:
3552     return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc);
3553   case Builtin::BI__builtin_coro_begin:
3554     return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin);
3555   case Builtin::BI__builtin_coro_end:
3556     return EmitCoroutineIntrinsic(E, Intrinsic::coro_end);
3557   case Builtin::BI__builtin_coro_suspend:
3558     return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend);
3559   case Builtin::BI__builtin_coro_param:
3560     return EmitCoroutineIntrinsic(E, Intrinsic::coro_param);
3561 
3562   // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions
3563   case Builtin::BIread_pipe:
3564   case Builtin::BIwrite_pipe: {
3565     Value *Arg0 = EmitScalarExpr(E->getArg(0)),
3566           *Arg1 = EmitScalarExpr(E->getArg(1));
3567     CGOpenCLRuntime OpenCLRT(CGM);
3568     Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3569     Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3570 
3571     // Type of the generic packet parameter.
3572     unsigned GenericAS =
3573         getContext().getTargetAddressSpace(LangAS::opencl_generic);
3574     llvm::Type *I8PTy = llvm::PointerType::get(
3575         llvm::Type::getInt8Ty(getLLVMContext()), GenericAS);
3576 
3577     // Testing which overloaded version we should generate the call for.
3578     if (2U == E->getNumArgs()) {
3579       const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2"
3580                                                              : "__write_pipe_2";
3581       // Creating a generic function type to be able to call with any builtin or
3582       // user defined type.
3583       llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty};
3584       llvm::FunctionType *FTy = llvm::FunctionType::get(
3585           Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3586       Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy);
3587       return RValue::get(
3588           Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3589                              {Arg0, BCast, PacketSize, PacketAlign}));
3590     } else {
3591       assert(4 == E->getNumArgs() &&
3592              "Illegal number of parameters to pipe function");
3593       const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4"
3594                                                              : "__write_pipe_4";
3595 
3596       llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy,
3597                               Int32Ty, Int32Ty};
3598       Value *Arg2 = EmitScalarExpr(E->getArg(2)),
3599             *Arg3 = EmitScalarExpr(E->getArg(3));
3600       llvm::FunctionType *FTy = llvm::FunctionType::get(
3601           Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3602       Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy);
3603       // We know the third argument is an integer type, but we may need to cast
3604       // it to i32.
3605       if (Arg2->getType() != Int32Ty)
3606         Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty);
3607       return RValue::get(Builder.CreateCall(
3608           CGM.CreateRuntimeFunction(FTy, Name),
3609           {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign}));
3610     }
3611   }
3612   // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write
3613   // functions
3614   case Builtin::BIreserve_read_pipe:
3615   case Builtin::BIreserve_write_pipe:
3616   case Builtin::BIwork_group_reserve_read_pipe:
3617   case Builtin::BIwork_group_reserve_write_pipe:
3618   case Builtin::BIsub_group_reserve_read_pipe:
3619   case Builtin::BIsub_group_reserve_write_pipe: {
3620     // Composing the mangled name for the function.
3621     const char *Name;
3622     if (BuiltinID == Builtin::BIreserve_read_pipe)
3623       Name = "__reserve_read_pipe";
3624     else if (BuiltinID == Builtin::BIreserve_write_pipe)
3625       Name = "__reserve_write_pipe";
3626     else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe)
3627       Name = "__work_group_reserve_read_pipe";
3628     else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe)
3629       Name = "__work_group_reserve_write_pipe";
3630     else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe)
3631       Name = "__sub_group_reserve_read_pipe";
3632     else
3633       Name = "__sub_group_reserve_write_pipe";
3634 
3635     Value *Arg0 = EmitScalarExpr(E->getArg(0)),
3636           *Arg1 = EmitScalarExpr(E->getArg(1));
3637     llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy);
3638     CGOpenCLRuntime OpenCLRT(CGM);
3639     Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3640     Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3641 
3642     // Building the generic function prototype.
3643     llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty};
3644     llvm::FunctionType *FTy = llvm::FunctionType::get(
3645         ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3646     // We know the second argument is an integer type, but we may need to cast
3647     // it to i32.
3648     if (Arg1->getType() != Int32Ty)
3649       Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty);
3650     return RValue::get(
3651         Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3652                            {Arg0, Arg1, PacketSize, PacketAlign}));
3653   }
3654   // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write
3655   // functions
3656   case Builtin::BIcommit_read_pipe:
3657   case Builtin::BIcommit_write_pipe:
3658   case Builtin::BIwork_group_commit_read_pipe:
3659   case Builtin::BIwork_group_commit_write_pipe:
3660   case Builtin::BIsub_group_commit_read_pipe:
3661   case Builtin::BIsub_group_commit_write_pipe: {
3662     const char *Name;
3663     if (BuiltinID == Builtin::BIcommit_read_pipe)
3664       Name = "__commit_read_pipe";
3665     else if (BuiltinID == Builtin::BIcommit_write_pipe)
3666       Name = "__commit_write_pipe";
3667     else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe)
3668       Name = "__work_group_commit_read_pipe";
3669     else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe)
3670       Name = "__work_group_commit_write_pipe";
3671     else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe)
3672       Name = "__sub_group_commit_read_pipe";
3673     else
3674       Name = "__sub_group_commit_write_pipe";
3675 
3676     Value *Arg0 = EmitScalarExpr(E->getArg(0)),
3677           *Arg1 = EmitScalarExpr(E->getArg(1));
3678     CGOpenCLRuntime OpenCLRT(CGM);
3679     Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3680     Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3681 
3682     // Building the generic function prototype.
3683     llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty};
3684     llvm::FunctionType *FTy =
3685         llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()),
3686                                 llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3687 
3688     return RValue::get(
3689         Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3690                            {Arg0, Arg1, PacketSize, PacketAlign}));
3691   }
3692   // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions
3693   case Builtin::BIget_pipe_num_packets:
3694   case Builtin::BIget_pipe_max_packets: {
3695     const char *BaseName;
3696     const auto *PipeTy = E->getArg(0)->getType()->castAs<PipeType>();
3697     if (BuiltinID == Builtin::BIget_pipe_num_packets)
3698       BaseName = "__get_pipe_num_packets";
3699     else
3700       BaseName = "__get_pipe_max_packets";
3701     std::string Name = std::string(BaseName) +
3702                        std::string(PipeTy->isReadOnly() ? "_ro" : "_wo");
3703 
3704     // Building the generic function prototype.
3705     Value *Arg0 = EmitScalarExpr(E->getArg(0));
3706     CGOpenCLRuntime OpenCLRT(CGM);
3707     Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3708     Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3709     llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty};
3710     llvm::FunctionType *FTy = llvm::FunctionType::get(
3711         Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3712 
3713     return RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3714                                           {Arg0, PacketSize, PacketAlign}));
3715   }
3716 
3717   // OpenCL v2.0 s6.13.9 - Address space qualifier functions.
3718   case Builtin::BIto_global:
3719   case Builtin::BIto_local:
3720   case Builtin::BIto_private: {
3721     auto Arg0 = EmitScalarExpr(E->getArg(0));
3722     auto NewArgT = llvm::PointerType::get(Int8Ty,
3723       CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
3724     auto NewRetT = llvm::PointerType::get(Int8Ty,
3725       CGM.getContext().getTargetAddressSpace(
3726         E->getType()->getPointeeType().getAddressSpace()));
3727     auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false);
3728     llvm::Value *NewArg;
3729     if (Arg0->getType()->getPointerAddressSpace() !=
3730         NewArgT->getPointerAddressSpace())
3731       NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT);
3732     else
3733       NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT);
3734     auto NewName = std::string("__") + E->getDirectCallee()->getName().str();
3735     auto NewCall =
3736         Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg});
3737     return RValue::get(Builder.CreateBitOrPointerCast(NewCall,
3738       ConvertType(E->getType())));
3739   }
3740 
3741   // OpenCL v2.0, s6.13.17 - Enqueue kernel function.
3742   // It contains four different overload formats specified in Table 6.13.17.1.
3743   case Builtin::BIenqueue_kernel: {
3744     StringRef Name; // Generated function call name
3745     unsigned NumArgs = E->getNumArgs();
3746 
3747     llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy);
3748     llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3749         getContext().getTargetAddressSpace(LangAS::opencl_generic));
3750 
3751     llvm::Value *Queue = EmitScalarExpr(E->getArg(0));
3752     llvm::Value *Flags = EmitScalarExpr(E->getArg(1));
3753     LValue NDRangeL = EmitAggExprToLValue(E->getArg(2));
3754     llvm::Value *Range = NDRangeL.getAddress(*this).getPointer();
3755     llvm::Type *RangeTy = NDRangeL.getAddress(*this).getType();
3756 
3757     if (NumArgs == 4) {
3758       // The most basic form of the call with parameters:
3759       // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void)
3760       Name = "__enqueue_kernel_basic";
3761       llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy,
3762                               GenericVoidPtrTy};
3763       llvm::FunctionType *FTy = llvm::FunctionType::get(
3764           Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3765 
3766       auto Info =
3767           CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
3768       llvm::Value *Kernel =
3769           Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3770       llvm::Value *Block =
3771           Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3772 
3773       AttrBuilder B;
3774       B.addByValAttr(NDRangeL.getAddress(*this).getElementType());
3775       llvm::AttributeList ByValAttrSet =
3776           llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B);
3777 
3778       auto RTCall =
3779           Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet),
3780                              {Queue, Flags, Range, Kernel, Block});
3781       RTCall->setAttributes(ByValAttrSet);
3782       return RValue::get(RTCall);
3783     }
3784     assert(NumArgs >= 5 && "Invalid enqueue_kernel signature");
3785 
3786     // Create a temporary array to hold the sizes of local pointer arguments
3787     // for the block. \p First is the position of the first size argument.
3788     auto CreateArrayForSizeVar = [=](unsigned First)
3789         -> std::tuple<llvm::Value *, llvm::Value *, llvm::Value *> {
3790       llvm::APInt ArraySize(32, NumArgs - First);
3791       QualType SizeArrayTy = getContext().getConstantArrayType(
3792           getContext().getSizeType(), ArraySize, nullptr, ArrayType::Normal,
3793           /*IndexTypeQuals=*/0);
3794       auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes");
3795       llvm::Value *TmpPtr = Tmp.getPointer();
3796       llvm::Value *TmpSize = EmitLifetimeStart(
3797           CGM.getDataLayout().getTypeAllocSize(Tmp.getElementType()), TmpPtr);
3798       llvm::Value *ElemPtr;
3799       // Each of the following arguments specifies the size of the corresponding
3800       // argument passed to the enqueued block.
3801       auto *Zero = llvm::ConstantInt::get(IntTy, 0);
3802       for (unsigned I = First; I < NumArgs; ++I) {
3803         auto *Index = llvm::ConstantInt::get(IntTy, I - First);
3804         auto *GEP = Builder.CreateGEP(TmpPtr, {Zero, Index});
3805         if (I == First)
3806           ElemPtr = GEP;
3807         auto *V =
3808             Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy);
3809         Builder.CreateAlignedStore(
3810             V, GEP, CGM.getDataLayout().getPrefTypeAlignment(SizeTy));
3811       }
3812       return std::tie(ElemPtr, TmpSize, TmpPtr);
3813     };
3814 
3815     // Could have events and/or varargs.
3816     if (E->getArg(3)->getType()->isBlockPointerType()) {
3817       // No events passed, but has variadic arguments.
3818       Name = "__enqueue_kernel_varargs";
3819       auto Info =
3820           CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
3821       llvm::Value *Kernel =
3822           Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3823       auto *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3824       llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
3825       std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(4);
3826 
3827       // Create a vector of the arguments, as well as a constant value to
3828       // express to the runtime the number of variadic arguments.
3829       std::vector<llvm::Value *> Args = {
3830           Queue,  Flags, Range,
3831           Kernel, Block, ConstantInt::get(IntTy, NumArgs - 4),
3832           ElemPtr};
3833       std::vector<llvm::Type *> ArgTys = {
3834           QueueTy,          IntTy, RangeTy,           GenericVoidPtrTy,
3835           GenericVoidPtrTy, IntTy, ElemPtr->getType()};
3836 
3837       llvm::FunctionType *FTy = llvm::FunctionType::get(
3838           Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3839       auto Call =
3840           RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3841                                          llvm::ArrayRef<llvm::Value *>(Args)));
3842       if (TmpSize)
3843         EmitLifetimeEnd(TmpSize, TmpPtr);
3844       return Call;
3845     }
3846     // Any calls now have event arguments passed.
3847     if (NumArgs >= 7) {
3848       llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy);
3849       llvm::PointerType *EventPtrTy = EventTy->getPointerTo(
3850           CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
3851 
3852       llvm::Value *NumEvents =
3853           Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty);
3854 
3855       // Since SemaOpenCLBuiltinEnqueueKernel allows fifth and sixth arguments
3856       // to be a null pointer constant (including `0` literal), we can take it
3857       // into account and emit null pointer directly.
3858       llvm::Value *EventWaitList = nullptr;
3859       if (E->getArg(4)->isNullPointerConstant(
3860               getContext(), Expr::NPC_ValueDependentIsNotNull)) {
3861         EventWaitList = llvm::ConstantPointerNull::get(EventPtrTy);
3862       } else {
3863         EventWaitList = E->getArg(4)->getType()->isArrayType()
3864                         ? EmitArrayToPointerDecay(E->getArg(4)).getPointer()
3865                         : EmitScalarExpr(E->getArg(4));
3866         // Convert to generic address space.
3867         EventWaitList = Builder.CreatePointerCast(EventWaitList, EventPtrTy);
3868       }
3869       llvm::Value *EventRet = nullptr;
3870       if (E->getArg(5)->isNullPointerConstant(
3871               getContext(), Expr::NPC_ValueDependentIsNotNull)) {
3872         EventRet = llvm::ConstantPointerNull::get(EventPtrTy);
3873       } else {
3874         EventRet =
3875             Builder.CreatePointerCast(EmitScalarExpr(E->getArg(5)), EventPtrTy);
3876       }
3877 
3878       auto Info =
3879           CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(6));
3880       llvm::Value *Kernel =
3881           Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3882       llvm::Value *Block =
3883           Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3884 
3885       std::vector<llvm::Type *> ArgTys = {
3886           QueueTy,    Int32Ty,    RangeTy,          Int32Ty,
3887           EventPtrTy, EventPtrTy, GenericVoidPtrTy, GenericVoidPtrTy};
3888 
3889       std::vector<llvm::Value *> Args = {Queue,     Flags,         Range,
3890                                          NumEvents, EventWaitList, EventRet,
3891                                          Kernel,    Block};
3892 
3893       if (NumArgs == 7) {
3894         // Has events but no variadics.
3895         Name = "__enqueue_kernel_basic_events";
3896         llvm::FunctionType *FTy = llvm::FunctionType::get(
3897             Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3898         return RValue::get(
3899             Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3900                                llvm::ArrayRef<llvm::Value *>(Args)));
3901       }
3902       // Has event info and variadics
3903       // Pass the number of variadics to the runtime function too.
3904       Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7));
3905       ArgTys.push_back(Int32Ty);
3906       Name = "__enqueue_kernel_events_varargs";
3907 
3908       llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
3909       std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(7);
3910       Args.push_back(ElemPtr);
3911       ArgTys.push_back(ElemPtr->getType());
3912 
3913       llvm::FunctionType *FTy = llvm::FunctionType::get(
3914           Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3915       auto Call =
3916           RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3917                                          llvm::ArrayRef<llvm::Value *>(Args)));
3918       if (TmpSize)
3919         EmitLifetimeEnd(TmpSize, TmpPtr);
3920       return Call;
3921     }
3922     LLVM_FALLTHROUGH;
3923   }
3924   // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block
3925   // parameter.
3926   case Builtin::BIget_kernel_work_group_size: {
3927     llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3928         getContext().getTargetAddressSpace(LangAS::opencl_generic));
3929     auto Info =
3930         CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
3931     Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3932     Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3933     return RValue::get(Builder.CreateCall(
3934         CGM.CreateRuntimeFunction(
3935             llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
3936                                     false),
3937             "__get_kernel_work_group_size_impl"),
3938         {Kernel, Arg}));
3939   }
3940   case Builtin::BIget_kernel_preferred_work_group_size_multiple: {
3941     llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3942         getContext().getTargetAddressSpace(LangAS::opencl_generic));
3943     auto Info =
3944         CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
3945     Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3946     Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3947     return RValue::get(Builder.CreateCall(
3948         CGM.CreateRuntimeFunction(
3949             llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
3950                                     false),
3951             "__get_kernel_preferred_work_group_size_multiple_impl"),
3952         {Kernel, Arg}));
3953   }
3954   case Builtin::BIget_kernel_max_sub_group_size_for_ndrange:
3955   case Builtin::BIget_kernel_sub_group_count_for_ndrange: {
3956     llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3957         getContext().getTargetAddressSpace(LangAS::opencl_generic));
3958     LValue NDRangeL = EmitAggExprToLValue(E->getArg(0));
3959     llvm::Value *NDRange = NDRangeL.getAddress(*this).getPointer();
3960     auto Info =
3961         CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(1));
3962     Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3963     Value *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3964     const char *Name =
3965         BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange
3966             ? "__get_kernel_max_sub_group_size_for_ndrange_impl"
3967             : "__get_kernel_sub_group_count_for_ndrange_impl";
3968     return RValue::get(Builder.CreateCall(
3969         CGM.CreateRuntimeFunction(
3970             llvm::FunctionType::get(
3971                 IntTy, {NDRange->getType(), GenericVoidPtrTy, GenericVoidPtrTy},
3972                 false),
3973             Name),
3974         {NDRange, Kernel, Block}));
3975   }
3976 
3977   case Builtin::BI__builtin_store_half:
3978   case Builtin::BI__builtin_store_halff: {
3979     Value *Val = EmitScalarExpr(E->getArg(0));
3980     Address Address = EmitPointerWithAlignment(E->getArg(1));
3981     Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy());
3982     return RValue::get(Builder.CreateStore(HalfVal, Address));
3983   }
3984   case Builtin::BI__builtin_load_half: {
3985     Address Address = EmitPointerWithAlignment(E->getArg(0));
3986     Value *HalfVal = Builder.CreateLoad(Address);
3987     return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy()));
3988   }
3989   case Builtin::BI__builtin_load_halff: {
3990     Address Address = EmitPointerWithAlignment(E->getArg(0));
3991     Value *HalfVal = Builder.CreateLoad(Address);
3992     return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy()));
3993   }
3994   case Builtin::BIprintf:
3995     if (getTarget().getTriple().isNVPTX())
3996       return EmitNVPTXDevicePrintfCallExpr(E, ReturnValue);
3997     break;
3998   case Builtin::BI__builtin_canonicalize:
3999   case Builtin::BI__builtin_canonicalizef:
4000   case Builtin::BI__builtin_canonicalizef16:
4001   case Builtin::BI__builtin_canonicalizel:
4002     return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize));
4003 
4004   case Builtin::BI__builtin_thread_pointer: {
4005     if (!getContext().getTargetInfo().isTLSSupported())
4006       CGM.ErrorUnsupported(E, "__builtin_thread_pointer");
4007     // Fall through - it's already mapped to the intrinsic by GCCBuiltin.
4008     break;
4009   }
4010   case Builtin::BI__builtin_os_log_format:
4011     return emitBuiltinOSLogFormat(*E);
4012 
4013   case Builtin::BI__xray_customevent: {
4014     if (!ShouldXRayInstrumentFunction())
4015       return RValue::getIgnored();
4016 
4017     if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
4018             XRayInstrKind::Custom))
4019       return RValue::getIgnored();
4020 
4021     if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
4022       if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayCustomEvents())
4023         return RValue::getIgnored();
4024 
4025     Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent);
4026     auto FTy = F->getFunctionType();
4027     auto Arg0 = E->getArg(0);
4028     auto Arg0Val = EmitScalarExpr(Arg0);
4029     auto Arg0Ty = Arg0->getType();
4030     auto PTy0 = FTy->getParamType(0);
4031     if (PTy0 != Arg0Val->getType()) {
4032       if (Arg0Ty->isArrayType())
4033         Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer();
4034       else
4035         Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0);
4036     }
4037     auto Arg1 = EmitScalarExpr(E->getArg(1));
4038     auto PTy1 = FTy->getParamType(1);
4039     if (PTy1 != Arg1->getType())
4040       Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1);
4041     return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1}));
4042   }
4043 
4044   case Builtin::BI__xray_typedevent: {
4045     // TODO: There should be a way to always emit events even if the current
4046     // function is not instrumented. Losing events in a stream can cripple
4047     // a trace.
4048     if (!ShouldXRayInstrumentFunction())
4049       return RValue::getIgnored();
4050 
4051     if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
4052             XRayInstrKind::Typed))
4053       return RValue::getIgnored();
4054 
4055     if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
4056       if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayTypedEvents())
4057         return RValue::getIgnored();
4058 
4059     Function *F = CGM.getIntrinsic(Intrinsic::xray_typedevent);
4060     auto FTy = F->getFunctionType();
4061     auto Arg0 = EmitScalarExpr(E->getArg(0));
4062     auto PTy0 = FTy->getParamType(0);
4063     if (PTy0 != Arg0->getType())
4064       Arg0 = Builder.CreateTruncOrBitCast(Arg0, PTy0);
4065     auto Arg1 = E->getArg(1);
4066     auto Arg1Val = EmitScalarExpr(Arg1);
4067     auto Arg1Ty = Arg1->getType();
4068     auto PTy1 = FTy->getParamType(1);
4069     if (PTy1 != Arg1Val->getType()) {
4070       if (Arg1Ty->isArrayType())
4071         Arg1Val = EmitArrayToPointerDecay(Arg1).getPointer();
4072       else
4073         Arg1Val = Builder.CreatePointerCast(Arg1Val, PTy1);
4074     }
4075     auto Arg2 = EmitScalarExpr(E->getArg(2));
4076     auto PTy2 = FTy->getParamType(2);
4077     if (PTy2 != Arg2->getType())
4078       Arg2 = Builder.CreateTruncOrBitCast(Arg2, PTy2);
4079     return RValue::get(Builder.CreateCall(F, {Arg0, Arg1Val, Arg2}));
4080   }
4081 
4082   case Builtin::BI__builtin_ms_va_start:
4083   case Builtin::BI__builtin_ms_va_end:
4084     return RValue::get(
4085         EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(),
4086                        BuiltinID == Builtin::BI__builtin_ms_va_start));
4087 
4088   case Builtin::BI__builtin_ms_va_copy: {
4089     // Lower this manually. We can't reliably determine whether or not any
4090     // given va_copy() is for a Win64 va_list from the calling convention
4091     // alone, because it's legal to do this from a System V ABI function.
4092     // With opaque pointer types, we won't have enough information in LLVM
4093     // IR to determine this from the argument types, either. Best to do it
4094     // now, while we have enough information.
4095     Address DestAddr = EmitMSVAListRef(E->getArg(0));
4096     Address SrcAddr = EmitMSVAListRef(E->getArg(1));
4097 
4098     llvm::Type *BPP = Int8PtrPtrTy;
4099 
4100     DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"),
4101                        DestAddr.getAlignment());
4102     SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"),
4103                       SrcAddr.getAlignment());
4104 
4105     Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val");
4106     return RValue::get(Builder.CreateStore(ArgPtr, DestAddr));
4107   }
4108   }
4109 
4110   // If this is an alias for a lib function (e.g. __builtin_sin), emit
4111   // the call using the normal call path, but using the unmangled
4112   // version of the function name.
4113   if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
4114     return emitLibraryCall(*this, FD, E,
4115                            CGM.getBuiltinLibFunction(FD, BuiltinID));
4116 
4117   // If this is a predefined lib function (e.g. malloc), emit the call
4118   // using exactly the normal call path.
4119   if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
4120     return emitLibraryCall(*this, FD, E,
4121                       cast<llvm::Constant>(EmitScalarExpr(E->getCallee())));
4122 
4123   // Check that a call to a target specific builtin has the correct target
4124   // features.
4125   // This is down here to avoid non-target specific builtins, however, if
4126   // generic builtins start to require generic target features then we
4127   // can move this up to the beginning of the function.
4128   checkTargetFeatures(E, FD);
4129 
4130   if (unsigned VectorWidth = getContext().BuiltinInfo.getRequiredVectorWidth(BuiltinID))
4131     LargestVectorWidth = std::max(LargestVectorWidth, VectorWidth);
4132 
4133   // See if we have a target specific intrinsic.
4134   const char *Name = getContext().BuiltinInfo.getName(BuiltinID);
4135   Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
4136   StringRef Prefix =
4137       llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch());
4138   if (!Prefix.empty()) {
4139     IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix.data(), Name);
4140     // NOTE we don't need to perform a compatibility flag check here since the
4141     // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the
4142     // MS builtins via ALL_MS_LANGUAGES and are filtered earlier.
4143     if (IntrinsicID == Intrinsic::not_intrinsic)
4144       IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name);
4145   }
4146 
4147   if (IntrinsicID != Intrinsic::not_intrinsic) {
4148     SmallVector<Value*, 16> Args;
4149 
4150     // Find out if any arguments are required to be integer constant
4151     // expressions.
4152     unsigned ICEArguments = 0;
4153     ASTContext::GetBuiltinTypeError Error;
4154     getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
4155     assert(Error == ASTContext::GE_None && "Should not codegen an error");
4156 
4157     Function *F = CGM.getIntrinsic(IntrinsicID);
4158     llvm::FunctionType *FTy = F->getFunctionType();
4159 
4160     for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
4161       Value *ArgValue;
4162       // If this is a normal argument, just emit it as a scalar.
4163       if ((ICEArguments & (1 << i)) == 0) {
4164         ArgValue = EmitScalarExpr(E->getArg(i));
4165       } else {
4166         // If this is required to be a constant, constant fold it so that we
4167         // know that the generated intrinsic gets a ConstantInt.
4168         llvm::APSInt Result;
4169         bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
4170         assert(IsConst && "Constant arg isn't actually constant?");
4171         (void)IsConst;
4172         ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
4173       }
4174 
4175       // If the intrinsic arg type is different from the builtin arg type
4176       // we need to do a bit cast.
4177       llvm::Type *PTy = FTy->getParamType(i);
4178       if (PTy != ArgValue->getType()) {
4179         // XXX - vector of pointers?
4180         if (auto *PtrTy = dyn_cast<llvm::PointerType>(PTy)) {
4181           if (PtrTy->getAddressSpace() !=
4182               ArgValue->getType()->getPointerAddressSpace()) {
4183             ArgValue = Builder.CreateAddrSpaceCast(
4184               ArgValue,
4185               ArgValue->getType()->getPointerTo(PtrTy->getAddressSpace()));
4186           }
4187         }
4188 
4189         assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
4190                "Must be able to losslessly bit cast to param");
4191         ArgValue = Builder.CreateBitCast(ArgValue, PTy);
4192       }
4193 
4194       Args.push_back(ArgValue);
4195     }
4196 
4197     Value *V = Builder.CreateCall(F, Args);
4198     QualType BuiltinRetType = E->getType();
4199 
4200     llvm::Type *RetTy = VoidTy;
4201     if (!BuiltinRetType->isVoidType())
4202       RetTy = ConvertType(BuiltinRetType);
4203 
4204     if (RetTy != V->getType()) {
4205       // XXX - vector of pointers?
4206       if (auto *PtrTy = dyn_cast<llvm::PointerType>(RetTy)) {
4207         if (PtrTy->getAddressSpace() != V->getType()->getPointerAddressSpace()) {
4208           V = Builder.CreateAddrSpaceCast(
4209             V, V->getType()->getPointerTo(PtrTy->getAddressSpace()));
4210         }
4211       }
4212 
4213       assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
4214              "Must be able to losslessly bit cast result type");
4215       V = Builder.CreateBitCast(V, RetTy);
4216     }
4217 
4218     return RValue::get(V);
4219   }
4220 
4221   // See if we have a target specific builtin that needs to be lowered.
4222   if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E, ReturnValue))
4223     return RValue::get(V);
4224 
4225   ErrorUnsupported(E, "builtin function");
4226 
4227   // Unknown builtin, for now just dump it out and return undef.
4228   return GetUndefRValue(E->getType());
4229 }
4230 
4231 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF,
4232                                         unsigned BuiltinID, const CallExpr *E,
4233                                         ReturnValueSlot ReturnValue,
4234                                         llvm::Triple::ArchType Arch) {
4235   switch (Arch) {
4236   case llvm::Triple::arm:
4237   case llvm::Triple::armeb:
4238   case llvm::Triple::thumb:
4239   case llvm::Triple::thumbeb:
4240     return CGF->EmitARMBuiltinExpr(BuiltinID, E, ReturnValue, Arch);
4241   case llvm::Triple::aarch64:
4242   case llvm::Triple::aarch64_32:
4243   case llvm::Triple::aarch64_be:
4244     return CGF->EmitAArch64BuiltinExpr(BuiltinID, E, Arch);
4245   case llvm::Triple::bpfeb:
4246   case llvm::Triple::bpfel:
4247     return CGF->EmitBPFBuiltinExpr(BuiltinID, E);
4248   case llvm::Triple::x86:
4249   case llvm::Triple::x86_64:
4250     return CGF->EmitX86BuiltinExpr(BuiltinID, E);
4251   case llvm::Triple::ppc:
4252   case llvm::Triple::ppc64:
4253   case llvm::Triple::ppc64le:
4254     return CGF->EmitPPCBuiltinExpr(BuiltinID, E);
4255   case llvm::Triple::r600:
4256   case llvm::Triple::amdgcn:
4257     return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E);
4258   case llvm::Triple::systemz:
4259     return CGF->EmitSystemZBuiltinExpr(BuiltinID, E);
4260   case llvm::Triple::nvptx:
4261   case llvm::Triple::nvptx64:
4262     return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E);
4263   case llvm::Triple::wasm32:
4264   case llvm::Triple::wasm64:
4265     return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E);
4266   case llvm::Triple::hexagon:
4267     return CGF->EmitHexagonBuiltinExpr(BuiltinID, E);
4268   default:
4269     return nullptr;
4270   }
4271 }
4272 
4273 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
4274                                               const CallExpr *E,
4275                                               ReturnValueSlot ReturnValue) {
4276   if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) {
4277     assert(getContext().getAuxTargetInfo() && "Missing aux target info");
4278     return EmitTargetArchBuiltinExpr(
4279         this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E,
4280         ReturnValue, getContext().getAuxTargetInfo()->getTriple().getArch());
4281   }
4282 
4283   return EmitTargetArchBuiltinExpr(this, BuiltinID, E, ReturnValue,
4284                                    getTarget().getTriple().getArch());
4285 }
4286 
4287 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
4288                                      NeonTypeFlags TypeFlags,
4289                                      bool HasLegalHalfType=true,
4290                                      bool V1Ty=false) {
4291   int IsQuad = TypeFlags.isQuad();
4292   switch (TypeFlags.getEltType()) {
4293   case NeonTypeFlags::Int8:
4294   case NeonTypeFlags::Poly8:
4295     return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
4296   case NeonTypeFlags::Int16:
4297   case NeonTypeFlags::Poly16:
4298     return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
4299   case NeonTypeFlags::Float16:
4300     if (HasLegalHalfType)
4301       return llvm::VectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad));
4302     else
4303       return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
4304   case NeonTypeFlags::Int32:
4305     return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
4306   case NeonTypeFlags::Int64:
4307   case NeonTypeFlags::Poly64:
4308     return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad));
4309   case NeonTypeFlags::Poly128:
4310     // FIXME: i128 and f128 doesn't get fully support in Clang and llvm.
4311     // There is a lot of i128 and f128 API missing.
4312     // so we use v16i8 to represent poly128 and get pattern matched.
4313     return llvm::VectorType::get(CGF->Int8Ty, 16);
4314   case NeonTypeFlags::Float32:
4315     return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad));
4316   case NeonTypeFlags::Float64:
4317     return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad));
4318   }
4319   llvm_unreachable("Unknown vector element type!");
4320 }
4321 
4322 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF,
4323                                           NeonTypeFlags IntTypeFlags) {
4324   int IsQuad = IntTypeFlags.isQuad();
4325   switch (IntTypeFlags.getEltType()) {
4326   case NeonTypeFlags::Int16:
4327     return llvm::VectorType::get(CGF->HalfTy, (4 << IsQuad));
4328   case NeonTypeFlags::Int32:
4329     return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad));
4330   case NeonTypeFlags::Int64:
4331     return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad));
4332   default:
4333     llvm_unreachable("Type can't be converted to floating-point!");
4334   }
4335 }
4336 
4337 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
4338   unsigned nElts = V->getType()->getVectorNumElements();
4339   Value* SV = llvm::ConstantVector::getSplat(nElts, C);
4340   return Builder.CreateShuffleVector(V, V, SV, "lane");
4341 }
4342 
4343 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
4344                                      const char *name,
4345                                      unsigned shift, bool rightshift) {
4346   unsigned j = 0;
4347   for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
4348        ai != ae; ++ai, ++j)
4349     if (shift > 0 && shift == j)
4350       Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
4351     else
4352       Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
4353 
4354   return Builder.CreateCall(F, Ops, name);
4355 }
4356 
4357 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
4358                                             bool neg) {
4359   int SV = cast<ConstantInt>(V)->getSExtValue();
4360   return ConstantInt::get(Ty, neg ? -SV : SV);
4361 }
4362 
4363 // Right-shift a vector by a constant.
4364 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift,
4365                                           llvm::Type *Ty, bool usgn,
4366                                           const char *name) {
4367   llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
4368 
4369   int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue();
4370   int EltSize = VTy->getScalarSizeInBits();
4371 
4372   Vec = Builder.CreateBitCast(Vec, Ty);
4373 
4374   // lshr/ashr are undefined when the shift amount is equal to the vector
4375   // element size.
4376   if (ShiftAmt == EltSize) {
4377     if (usgn) {
4378       // Right-shifting an unsigned value by its size yields 0.
4379       return llvm::ConstantAggregateZero::get(VTy);
4380     } else {
4381       // Right-shifting a signed value by its size is equivalent
4382       // to a shift of size-1.
4383       --ShiftAmt;
4384       Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt);
4385     }
4386   }
4387 
4388   Shift = EmitNeonShiftVector(Shift, Ty, false);
4389   if (usgn)
4390     return Builder.CreateLShr(Vec, Shift, name);
4391   else
4392     return Builder.CreateAShr(Vec, Shift, name);
4393 }
4394 
4395 enum {
4396   AddRetType = (1 << 0),
4397   Add1ArgType = (1 << 1),
4398   Add2ArgTypes = (1 << 2),
4399 
4400   VectorizeRetType = (1 << 3),
4401   VectorizeArgTypes = (1 << 4),
4402 
4403   InventFloatType = (1 << 5),
4404   UnsignedAlts = (1 << 6),
4405 
4406   Use64BitVectors = (1 << 7),
4407   Use128BitVectors = (1 << 8),
4408 
4409   Vectorize1ArgType = Add1ArgType | VectorizeArgTypes,
4410   VectorRet = AddRetType | VectorizeRetType,
4411   VectorRetGetArgs01 =
4412       AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes,
4413   FpCmpzModifiers =
4414       AddRetType | VectorizeRetType | Add1ArgType | InventFloatType
4415 };
4416 
4417 namespace {
4418 struct NeonIntrinsicInfo {
4419   const char *NameHint;
4420   unsigned BuiltinID;
4421   unsigned LLVMIntrinsic;
4422   unsigned AltLLVMIntrinsic;
4423   unsigned TypeModifier;
4424 
4425   bool operator<(unsigned RHSBuiltinID) const {
4426     return BuiltinID < RHSBuiltinID;
4427   }
4428   bool operator<(const NeonIntrinsicInfo &TE) const {
4429     return BuiltinID < TE.BuiltinID;
4430   }
4431 };
4432 } // end anonymous namespace
4433 
4434 #define NEONMAP0(NameBase) \
4435   { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 }
4436 
4437 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
4438   { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
4439       Intrinsic::LLVMIntrinsic, 0, TypeModifier }
4440 
4441 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \
4442   { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
4443       Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \
4444       TypeModifier }
4445 
4446 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = {
4447   NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
4448   NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
4449   NEONMAP1(vabs_v, arm_neon_vabs, 0),
4450   NEONMAP1(vabsq_v, arm_neon_vabs, 0),
4451   NEONMAP0(vaddhn_v),
4452   NEONMAP1(vaesdq_v, arm_neon_aesd, 0),
4453   NEONMAP1(vaeseq_v, arm_neon_aese, 0),
4454   NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0),
4455   NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0),
4456   NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType),
4457   NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType),
4458   NEONMAP1(vcadd_rot270_v, arm_neon_vcadd_rot270, Add1ArgType),
4459   NEONMAP1(vcadd_rot90_v, arm_neon_vcadd_rot90, Add1ArgType),
4460   NEONMAP1(vcaddq_rot270_v, arm_neon_vcadd_rot270, Add1ArgType),
4461   NEONMAP1(vcaddq_rot90_v, arm_neon_vcadd_rot90, Add1ArgType),
4462   NEONMAP1(vcage_v, arm_neon_vacge, 0),
4463   NEONMAP1(vcageq_v, arm_neon_vacge, 0),
4464   NEONMAP1(vcagt_v, arm_neon_vacgt, 0),
4465   NEONMAP1(vcagtq_v, arm_neon_vacgt, 0),
4466   NEONMAP1(vcale_v, arm_neon_vacge, 0),
4467   NEONMAP1(vcaleq_v, arm_neon_vacge, 0),
4468   NEONMAP1(vcalt_v, arm_neon_vacgt, 0),
4469   NEONMAP1(vcaltq_v, arm_neon_vacgt, 0),
4470   NEONMAP0(vceqz_v),
4471   NEONMAP0(vceqzq_v),
4472   NEONMAP0(vcgez_v),
4473   NEONMAP0(vcgezq_v),
4474   NEONMAP0(vcgtz_v),
4475   NEONMAP0(vcgtzq_v),
4476   NEONMAP0(vclez_v),
4477   NEONMAP0(vclezq_v),
4478   NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType),
4479   NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType),
4480   NEONMAP0(vcltz_v),
4481   NEONMAP0(vcltzq_v),
4482   NEONMAP1(vclz_v, ctlz, Add1ArgType),
4483   NEONMAP1(vclzq_v, ctlz, Add1ArgType),
4484   NEONMAP1(vcnt_v, ctpop, Add1ArgType),
4485   NEONMAP1(vcntq_v, ctpop, Add1ArgType),
4486   NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0),
4487   NEONMAP0(vcvt_f16_v),
4488   NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
4489   NEONMAP0(vcvt_f32_v),
4490   NEONMAP2(vcvt_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4491   NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4492   NEONMAP1(vcvt_n_s16_v, arm_neon_vcvtfp2fxs, 0),
4493   NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
4494   NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
4495   NEONMAP1(vcvt_n_u16_v, arm_neon_vcvtfp2fxu, 0),
4496   NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
4497   NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
4498   NEONMAP0(vcvt_s16_v),
4499   NEONMAP0(vcvt_s32_v),
4500   NEONMAP0(vcvt_s64_v),
4501   NEONMAP0(vcvt_u16_v),
4502   NEONMAP0(vcvt_u32_v),
4503   NEONMAP0(vcvt_u64_v),
4504   NEONMAP1(vcvta_s16_v, arm_neon_vcvtas, 0),
4505   NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
4506   NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
4507   NEONMAP1(vcvta_u16_v, arm_neon_vcvtau, 0),
4508   NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
4509   NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
4510   NEONMAP1(vcvtaq_s16_v, arm_neon_vcvtas, 0),
4511   NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
4512   NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
4513   NEONMAP1(vcvtaq_u16_v, arm_neon_vcvtau, 0),
4514   NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
4515   NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
4516   NEONMAP1(vcvtm_s16_v, arm_neon_vcvtms, 0),
4517   NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
4518   NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
4519   NEONMAP1(vcvtm_u16_v, arm_neon_vcvtmu, 0),
4520   NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
4521   NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
4522   NEONMAP1(vcvtmq_s16_v, arm_neon_vcvtms, 0),
4523   NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
4524   NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
4525   NEONMAP1(vcvtmq_u16_v, arm_neon_vcvtmu, 0),
4526   NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
4527   NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
4528   NEONMAP1(vcvtn_s16_v, arm_neon_vcvtns, 0),
4529   NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
4530   NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
4531   NEONMAP1(vcvtn_u16_v, arm_neon_vcvtnu, 0),
4532   NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
4533   NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
4534   NEONMAP1(vcvtnq_s16_v, arm_neon_vcvtns, 0),
4535   NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
4536   NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
4537   NEONMAP1(vcvtnq_u16_v, arm_neon_vcvtnu, 0),
4538   NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
4539   NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
4540   NEONMAP1(vcvtp_s16_v, arm_neon_vcvtps, 0),
4541   NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
4542   NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
4543   NEONMAP1(vcvtp_u16_v, arm_neon_vcvtpu, 0),
4544   NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
4545   NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
4546   NEONMAP1(vcvtpq_s16_v, arm_neon_vcvtps, 0),
4547   NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
4548   NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
4549   NEONMAP1(vcvtpq_u16_v, arm_neon_vcvtpu, 0),
4550   NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
4551   NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
4552   NEONMAP0(vcvtq_f16_v),
4553   NEONMAP0(vcvtq_f32_v),
4554   NEONMAP2(vcvtq_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4555   NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4556   NEONMAP1(vcvtq_n_s16_v, arm_neon_vcvtfp2fxs, 0),
4557   NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
4558   NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
4559   NEONMAP1(vcvtq_n_u16_v, arm_neon_vcvtfp2fxu, 0),
4560   NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
4561   NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
4562   NEONMAP0(vcvtq_s16_v),
4563   NEONMAP0(vcvtq_s32_v),
4564   NEONMAP0(vcvtq_s64_v),
4565   NEONMAP0(vcvtq_u16_v),
4566   NEONMAP0(vcvtq_u32_v),
4567   NEONMAP0(vcvtq_u64_v),
4568   NEONMAP2(vdot_v, arm_neon_udot, arm_neon_sdot, 0),
4569   NEONMAP2(vdotq_v, arm_neon_udot, arm_neon_sdot, 0),
4570   NEONMAP0(vext_v),
4571   NEONMAP0(vextq_v),
4572   NEONMAP0(vfma_v),
4573   NEONMAP0(vfmaq_v),
4574   NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
4575   NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
4576   NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
4577   NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
4578   NEONMAP0(vld1_dup_v),
4579   NEONMAP1(vld1_v, arm_neon_vld1, 0),
4580   NEONMAP1(vld1_x2_v, arm_neon_vld1x2, 0),
4581   NEONMAP1(vld1_x3_v, arm_neon_vld1x3, 0),
4582   NEONMAP1(vld1_x4_v, arm_neon_vld1x4, 0),
4583   NEONMAP0(vld1q_dup_v),
4584   NEONMAP1(vld1q_v, arm_neon_vld1, 0),
4585   NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0),
4586   NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0),
4587   NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0),
4588   NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0),
4589   NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
4590   NEONMAP1(vld2_v, arm_neon_vld2, 0),
4591   NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0),
4592   NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
4593   NEONMAP1(vld2q_v, arm_neon_vld2, 0),
4594   NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0),
4595   NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
4596   NEONMAP1(vld3_v, arm_neon_vld3, 0),
4597   NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0),
4598   NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
4599   NEONMAP1(vld3q_v, arm_neon_vld3, 0),
4600   NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0),
4601   NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
4602   NEONMAP1(vld4_v, arm_neon_vld4, 0),
4603   NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0),
4604   NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
4605   NEONMAP1(vld4q_v, arm_neon_vld4, 0),
4606   NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
4607   NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType),
4608   NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType),
4609   NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
4610   NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
4611   NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType),
4612   NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType),
4613   NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
4614   NEONMAP0(vmovl_v),
4615   NEONMAP0(vmovn_v),
4616   NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType),
4617   NEONMAP0(vmull_v),
4618   NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType),
4619   NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
4620   NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
4621   NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType),
4622   NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
4623   NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
4624   NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType),
4625   NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts),
4626   NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts),
4627   NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType),
4628   NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType),
4629   NEONMAP2(vqadd_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts),
4630   NEONMAP2(vqaddq_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts),
4631   NEONMAP2(vqdmlal_v, arm_neon_vqdmull, sadd_sat, 0),
4632   NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, ssub_sat, 0),
4633   NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType),
4634   NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType),
4635   NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType),
4636   NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts),
4637   NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType),
4638   NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType),
4639   NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType),
4640   NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType),
4641   NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType),
4642   NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
4643   NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
4644   NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
4645   NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
4646   NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
4647   NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
4648   NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0),
4649   NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0),
4650   NEONMAP2(vqsub_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts),
4651   NEONMAP2(vqsubq_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts),
4652   NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType),
4653   NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
4654   NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
4655   NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType),
4656   NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType),
4657   NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
4658   NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
4659   NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType),
4660   NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType),
4661   NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType),
4662   NEONMAP0(vrndi_v),
4663   NEONMAP0(vrndiq_v),
4664   NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType),
4665   NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType),
4666   NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType),
4667   NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType),
4668   NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType),
4669   NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType),
4670   NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType),
4671   NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType),
4672   NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType),
4673   NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
4674   NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
4675   NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
4676   NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
4677   NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
4678   NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
4679   NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType),
4680   NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType),
4681   NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType),
4682   NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0),
4683   NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0),
4684   NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0),
4685   NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0),
4686   NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0),
4687   NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0),
4688   NEONMAP0(vshl_n_v),
4689   NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
4690   NEONMAP0(vshll_n_v),
4691   NEONMAP0(vshlq_n_v),
4692   NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
4693   NEONMAP0(vshr_n_v),
4694   NEONMAP0(vshrn_n_v),
4695   NEONMAP0(vshrq_n_v),
4696   NEONMAP1(vst1_v, arm_neon_vst1, 0),
4697   NEONMAP1(vst1_x2_v, arm_neon_vst1x2, 0),
4698   NEONMAP1(vst1_x3_v, arm_neon_vst1x3, 0),
4699   NEONMAP1(vst1_x4_v, arm_neon_vst1x4, 0),
4700   NEONMAP1(vst1q_v, arm_neon_vst1, 0),
4701   NEONMAP1(vst1q_x2_v, arm_neon_vst1x2, 0),
4702   NEONMAP1(vst1q_x3_v, arm_neon_vst1x3, 0),
4703   NEONMAP1(vst1q_x4_v, arm_neon_vst1x4, 0),
4704   NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0),
4705   NEONMAP1(vst2_v, arm_neon_vst2, 0),
4706   NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0),
4707   NEONMAP1(vst2q_v, arm_neon_vst2, 0),
4708   NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0),
4709   NEONMAP1(vst3_v, arm_neon_vst3, 0),
4710   NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0),
4711   NEONMAP1(vst3q_v, arm_neon_vst3, 0),
4712   NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0),
4713   NEONMAP1(vst4_v, arm_neon_vst4, 0),
4714   NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0),
4715   NEONMAP1(vst4q_v, arm_neon_vst4, 0),
4716   NEONMAP0(vsubhn_v),
4717   NEONMAP0(vtrn_v),
4718   NEONMAP0(vtrnq_v),
4719   NEONMAP0(vtst_v),
4720   NEONMAP0(vtstq_v),
4721   NEONMAP0(vuzp_v),
4722   NEONMAP0(vuzpq_v),
4723   NEONMAP0(vzip_v),
4724   NEONMAP0(vzipq_v)
4725 };
4726 
4727 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
4728   NEONMAP1(vabs_v, aarch64_neon_abs, 0),
4729   NEONMAP1(vabsq_v, aarch64_neon_abs, 0),
4730   NEONMAP0(vaddhn_v),
4731   NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0),
4732   NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0),
4733   NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0),
4734   NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0),
4735   NEONMAP1(vcadd_rot270_v, aarch64_neon_vcadd_rot270, Add1ArgType),
4736   NEONMAP1(vcadd_rot90_v, aarch64_neon_vcadd_rot90, Add1ArgType),
4737   NEONMAP1(vcaddq_rot270_v, aarch64_neon_vcadd_rot270, Add1ArgType),
4738   NEONMAP1(vcaddq_rot90_v, aarch64_neon_vcadd_rot90, Add1ArgType),
4739   NEONMAP1(vcage_v, aarch64_neon_facge, 0),
4740   NEONMAP1(vcageq_v, aarch64_neon_facge, 0),
4741   NEONMAP1(vcagt_v, aarch64_neon_facgt, 0),
4742   NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0),
4743   NEONMAP1(vcale_v, aarch64_neon_facge, 0),
4744   NEONMAP1(vcaleq_v, aarch64_neon_facge, 0),
4745   NEONMAP1(vcalt_v, aarch64_neon_facgt, 0),
4746   NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0),
4747   NEONMAP0(vceqz_v),
4748   NEONMAP0(vceqzq_v),
4749   NEONMAP0(vcgez_v),
4750   NEONMAP0(vcgezq_v),
4751   NEONMAP0(vcgtz_v),
4752   NEONMAP0(vcgtzq_v),
4753   NEONMAP0(vclez_v),
4754   NEONMAP0(vclezq_v),
4755   NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType),
4756   NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType),
4757   NEONMAP0(vcltz_v),
4758   NEONMAP0(vcltzq_v),
4759   NEONMAP1(vclz_v, ctlz, Add1ArgType),
4760   NEONMAP1(vclzq_v, ctlz, Add1ArgType),
4761   NEONMAP1(vcnt_v, ctpop, Add1ArgType),
4762   NEONMAP1(vcntq_v, ctpop, Add1ArgType),
4763   NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0),
4764   NEONMAP0(vcvt_f16_v),
4765   NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
4766   NEONMAP0(vcvt_f32_v),
4767   NEONMAP2(vcvt_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4768   NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4769   NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4770   NEONMAP1(vcvt_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
4771   NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
4772   NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
4773   NEONMAP1(vcvt_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
4774   NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
4775   NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
4776   NEONMAP0(vcvtq_f16_v),
4777   NEONMAP0(vcvtq_f32_v),
4778   NEONMAP2(vcvtq_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4779   NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4780   NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4781   NEONMAP1(vcvtq_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
4782   NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
4783   NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
4784   NEONMAP1(vcvtq_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
4785   NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
4786   NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
4787   NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
4788   NEONMAP2(vdot_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
4789   NEONMAP2(vdotq_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
4790   NEONMAP0(vext_v),
4791   NEONMAP0(vextq_v),
4792   NEONMAP0(vfma_v),
4793   NEONMAP0(vfmaq_v),
4794   NEONMAP1(vfmlal_high_v, aarch64_neon_fmlal2, 0),
4795   NEONMAP1(vfmlal_low_v, aarch64_neon_fmlal, 0),
4796   NEONMAP1(vfmlalq_high_v, aarch64_neon_fmlal2, 0),
4797   NEONMAP1(vfmlalq_low_v, aarch64_neon_fmlal, 0),
4798   NEONMAP1(vfmlsl_high_v, aarch64_neon_fmlsl2, 0),
4799   NEONMAP1(vfmlsl_low_v, aarch64_neon_fmlsl, 0),
4800   NEONMAP1(vfmlslq_high_v, aarch64_neon_fmlsl2, 0),
4801   NEONMAP1(vfmlslq_low_v, aarch64_neon_fmlsl, 0),
4802   NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
4803   NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
4804   NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
4805   NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
4806   NEONMAP1(vld1_x2_v, aarch64_neon_ld1x2, 0),
4807   NEONMAP1(vld1_x3_v, aarch64_neon_ld1x3, 0),
4808   NEONMAP1(vld1_x4_v, aarch64_neon_ld1x4, 0),
4809   NEONMAP1(vld1q_x2_v, aarch64_neon_ld1x2, 0),
4810   NEONMAP1(vld1q_x3_v, aarch64_neon_ld1x3, 0),
4811   NEONMAP1(vld1q_x4_v, aarch64_neon_ld1x4, 0),
4812   NEONMAP0(vmovl_v),
4813   NEONMAP0(vmovn_v),
4814   NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType),
4815   NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType),
4816   NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType),
4817   NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
4818   NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
4819   NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType),
4820   NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType),
4821   NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType),
4822   NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
4823   NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
4824   NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0),
4825   NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0),
4826   NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType),
4827   NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType),
4828   NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType),
4829   NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts),
4830   NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType),
4831   NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType),
4832   NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType),
4833   NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType),
4834   NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType),
4835   NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
4836   NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
4837   NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts),
4838   NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
4839   NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts),
4840   NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
4841   NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0),
4842   NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0),
4843   NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
4844   NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
4845   NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType),
4846   NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
4847   NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
4848   NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType),
4849   NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType),
4850   NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
4851   NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
4852   NEONMAP0(vrndi_v),
4853   NEONMAP0(vrndiq_v),
4854   NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
4855   NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
4856   NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
4857   NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
4858   NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
4859   NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
4860   NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType),
4861   NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType),
4862   NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType),
4863   NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0),
4864   NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0),
4865   NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0),
4866   NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0),
4867   NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0),
4868   NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0),
4869   NEONMAP0(vshl_n_v),
4870   NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
4871   NEONMAP0(vshll_n_v),
4872   NEONMAP0(vshlq_n_v),
4873   NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
4874   NEONMAP0(vshr_n_v),
4875   NEONMAP0(vshrn_n_v),
4876   NEONMAP0(vshrq_n_v),
4877   NEONMAP1(vst1_x2_v, aarch64_neon_st1x2, 0),
4878   NEONMAP1(vst1_x3_v, aarch64_neon_st1x3, 0),
4879   NEONMAP1(vst1_x4_v, aarch64_neon_st1x4, 0),
4880   NEONMAP1(vst1q_x2_v, aarch64_neon_st1x2, 0),
4881   NEONMAP1(vst1q_x3_v, aarch64_neon_st1x3, 0),
4882   NEONMAP1(vst1q_x4_v, aarch64_neon_st1x4, 0),
4883   NEONMAP0(vsubhn_v),
4884   NEONMAP0(vtst_v),
4885   NEONMAP0(vtstq_v),
4886 };
4887 
4888 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = {
4889   NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType),
4890   NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType),
4891   NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType),
4892   NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
4893   NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
4894   NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
4895   NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
4896   NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
4897   NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
4898   NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4899   NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
4900   NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType),
4901   NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
4902   NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType),
4903   NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4904   NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4905   NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
4906   NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
4907   NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
4908   NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
4909   NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
4910   NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
4911   NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
4912   NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
4913   NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
4914   NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
4915   NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
4916   NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
4917   NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
4918   NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
4919   NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
4920   NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
4921   NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
4922   NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
4923   NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
4924   NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
4925   NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
4926   NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
4927   NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
4928   NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
4929   NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
4930   NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
4931   NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
4932   NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
4933   NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
4934   NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
4935   NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
4936   NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
4937   NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0),
4938   NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4939   NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4940   NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4941   NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4942   NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
4943   NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
4944   NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4945   NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4946   NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
4947   NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
4948   NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4949   NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4950   NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4951   NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
4952   NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
4953   NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
4954   NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
4955   NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
4956   NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
4957   NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
4958   NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0),
4959   NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType),
4960   NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType),
4961   NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4962   NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4963   NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4964   NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4965   NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4966   NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4967   NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4968   NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4969   NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
4970   NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
4971   NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
4972   NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType),
4973   NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
4974   NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType),
4975   NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
4976   NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
4977   NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType),
4978   NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType),
4979   NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
4980   NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
4981   NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType),
4982   NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType),
4983   NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors),
4984   NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType),
4985   NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors),
4986   NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0),
4987   NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType),
4988   NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType),
4989   NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
4990   NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
4991   NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
4992   NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
4993   NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType),
4994   NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
4995   NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
4996   NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
4997   NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType),
4998   NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
4999   NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType),
5000   NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors),
5001   NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType),
5002   NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
5003   NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
5004   NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType),
5005   NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType),
5006   NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
5007   NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
5008   NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType),
5009   NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType),
5010   NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType),
5011   NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType),
5012   NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
5013   NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
5014   NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
5015   NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
5016   NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType),
5017   NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
5018   NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
5019   NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
5020   NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
5021   NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
5022   NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
5023   NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType),
5024   NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType),
5025   NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
5026   NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
5027   NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
5028   NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
5029   NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType),
5030   NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType),
5031   NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType),
5032   NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType),
5033   NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
5034   NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
5035   NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType),
5036   NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType),
5037   NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType),
5038   NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
5039   NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
5040   NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
5041   NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
5042   NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType),
5043   NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
5044   NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
5045   NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
5046   NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
5047   NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType),
5048   NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType),
5049   NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
5050   NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
5051   NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType),
5052   NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType),
5053   NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType),
5054   NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType),
5055   NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType),
5056   NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType),
5057   NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType),
5058   NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType),
5059   NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType),
5060   NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType),
5061   NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType),
5062   NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType),
5063   NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0),
5064   NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0),
5065   NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0),
5066   NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0),
5067   NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType),
5068   NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType),
5069   NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType),
5070   NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType),
5071   NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
5072   NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType),
5073   NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
5074   NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType),
5075   NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType),
5076   NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType),
5077   NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
5078   NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType),
5079   NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
5080   NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType),
5081   // FP16 scalar intrinisics go here.
5082   NEONMAP1(vabdh_f16, aarch64_sisd_fabd, Add1ArgType),
5083   NEONMAP1(vcvtah_s32_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
5084   NEONMAP1(vcvtah_s64_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
5085   NEONMAP1(vcvtah_u32_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
5086   NEONMAP1(vcvtah_u64_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
5087   NEONMAP1(vcvth_n_f16_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
5088   NEONMAP1(vcvth_n_f16_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
5089   NEONMAP1(vcvth_n_f16_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
5090   NEONMAP1(vcvth_n_f16_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
5091   NEONMAP1(vcvth_n_s32_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
5092   NEONMAP1(vcvth_n_s64_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
5093   NEONMAP1(vcvth_n_u32_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
5094   NEONMAP1(vcvth_n_u64_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
5095   NEONMAP1(vcvtmh_s32_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
5096   NEONMAP1(vcvtmh_s64_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
5097   NEONMAP1(vcvtmh_u32_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
5098   NEONMAP1(vcvtmh_u64_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
5099   NEONMAP1(vcvtnh_s32_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
5100   NEONMAP1(vcvtnh_s64_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
5101   NEONMAP1(vcvtnh_u32_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
5102   NEONMAP1(vcvtnh_u64_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
5103   NEONMAP1(vcvtph_s32_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
5104   NEONMAP1(vcvtph_s64_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
5105   NEONMAP1(vcvtph_u32_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
5106   NEONMAP1(vcvtph_u64_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
5107   NEONMAP1(vmulxh_f16, aarch64_neon_fmulx, Add1ArgType),
5108   NEONMAP1(vrecpeh_f16, aarch64_neon_frecpe, Add1ArgType),
5109   NEONMAP1(vrecpxh_f16, aarch64_neon_frecpx, Add1ArgType),
5110   NEONMAP1(vrsqrteh_f16, aarch64_neon_frsqrte, Add1ArgType),
5111   NEONMAP1(vrsqrtsh_f16, aarch64_neon_frsqrts, Add1ArgType),
5112 };
5113 
5114 #undef NEONMAP0
5115 #undef NEONMAP1
5116 #undef NEONMAP2
5117 
5118 static bool NEONSIMDIntrinsicsProvenSorted = false;
5119 
5120 static bool AArch64SIMDIntrinsicsProvenSorted = false;
5121 static bool AArch64SISDIntrinsicsProvenSorted = false;
5122 
5123 
5124 static const NeonIntrinsicInfo *
5125 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap,
5126                        unsigned BuiltinID, bool &MapProvenSorted) {
5127 
5128 #ifndef NDEBUG
5129   if (!MapProvenSorted) {
5130     assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap)));
5131     MapProvenSorted = true;
5132   }
5133 #endif
5134 
5135   const NeonIntrinsicInfo *Builtin = llvm::lower_bound(IntrinsicMap, BuiltinID);
5136 
5137   if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID)
5138     return Builtin;
5139 
5140   return nullptr;
5141 }
5142 
5143 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID,
5144                                                    unsigned Modifier,
5145                                                    llvm::Type *ArgType,
5146                                                    const CallExpr *E) {
5147   int VectorSize = 0;
5148   if (Modifier & Use64BitVectors)
5149     VectorSize = 64;
5150   else if (Modifier & Use128BitVectors)
5151     VectorSize = 128;
5152 
5153   // Return type.
5154   SmallVector<llvm::Type *, 3> Tys;
5155   if (Modifier & AddRetType) {
5156     llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
5157     if (Modifier & VectorizeRetType)
5158       Ty = llvm::VectorType::get(
5159           Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1);
5160 
5161     Tys.push_back(Ty);
5162   }
5163 
5164   // Arguments.
5165   if (Modifier & VectorizeArgTypes) {
5166     int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1;
5167     ArgType = llvm::VectorType::get(ArgType, Elts);
5168   }
5169 
5170   if (Modifier & (Add1ArgType | Add2ArgTypes))
5171     Tys.push_back(ArgType);
5172 
5173   if (Modifier & Add2ArgTypes)
5174     Tys.push_back(ArgType);
5175 
5176   if (Modifier & InventFloatType)
5177     Tys.push_back(FloatTy);
5178 
5179   return CGM.getIntrinsic(IntrinsicID, Tys);
5180 }
5181 
5182 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF,
5183                                             const NeonIntrinsicInfo &SISDInfo,
5184                                             SmallVectorImpl<Value *> &Ops,
5185                                             const CallExpr *E) {
5186   unsigned BuiltinID = SISDInfo.BuiltinID;
5187   unsigned int Int = SISDInfo.LLVMIntrinsic;
5188   unsigned Modifier = SISDInfo.TypeModifier;
5189   const char *s = SISDInfo.NameHint;
5190 
5191   switch (BuiltinID) {
5192   case NEON::BI__builtin_neon_vcled_s64:
5193   case NEON::BI__builtin_neon_vcled_u64:
5194   case NEON::BI__builtin_neon_vcles_f32:
5195   case NEON::BI__builtin_neon_vcled_f64:
5196   case NEON::BI__builtin_neon_vcltd_s64:
5197   case NEON::BI__builtin_neon_vcltd_u64:
5198   case NEON::BI__builtin_neon_vclts_f32:
5199   case NEON::BI__builtin_neon_vcltd_f64:
5200   case NEON::BI__builtin_neon_vcales_f32:
5201   case NEON::BI__builtin_neon_vcaled_f64:
5202   case NEON::BI__builtin_neon_vcalts_f32:
5203   case NEON::BI__builtin_neon_vcaltd_f64:
5204     // Only one direction of comparisons actually exist, cmle is actually a cmge
5205     // with swapped operands. The table gives us the right intrinsic but we
5206     // still need to do the swap.
5207     std::swap(Ops[0], Ops[1]);
5208     break;
5209   }
5210 
5211   assert(Int && "Generic code assumes a valid intrinsic");
5212 
5213   // Determine the type(s) of this overloaded AArch64 intrinsic.
5214   const Expr *Arg = E->getArg(0);
5215   llvm::Type *ArgTy = CGF.ConvertType(Arg->getType());
5216   Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E);
5217 
5218   int j = 0;
5219   ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0);
5220   for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
5221        ai != ae; ++ai, ++j) {
5222     llvm::Type *ArgTy = ai->getType();
5223     if (Ops[j]->getType()->getPrimitiveSizeInBits() ==
5224              ArgTy->getPrimitiveSizeInBits())
5225       continue;
5226 
5227     assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy());
5228     // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate
5229     // it before inserting.
5230     Ops[j] =
5231         CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType());
5232     Ops[j] =
5233         CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0);
5234   }
5235 
5236   Value *Result = CGF.EmitNeonCall(F, Ops, s);
5237   llvm::Type *ResultType = CGF.ConvertType(E->getType());
5238   if (ResultType->getPrimitiveSizeInBits() <
5239       Result->getType()->getPrimitiveSizeInBits())
5240     return CGF.Builder.CreateExtractElement(Result, C0);
5241 
5242   return CGF.Builder.CreateBitCast(Result, ResultType, s);
5243 }
5244 
5245 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
5246     unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic,
5247     const char *NameHint, unsigned Modifier, const CallExpr *E,
5248     SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1,
5249     llvm::Triple::ArchType Arch) {
5250   // Get the last argument, which specifies the vector type.
5251   llvm::APSInt NeonTypeConst;
5252   const Expr *Arg = E->getArg(E->getNumArgs() - 1);
5253   if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext()))
5254     return nullptr;
5255 
5256   // Determine the type of this overloaded NEON intrinsic.
5257   NeonTypeFlags Type(NeonTypeConst.getZExtValue());
5258   bool Usgn = Type.isUnsigned();
5259   bool Quad = Type.isQuad();
5260   const bool HasLegalHalfType = getTarget().hasLegalHalfType();
5261 
5262   llvm::VectorType *VTy = GetNeonType(this, Type, HasLegalHalfType);
5263   llvm::Type *Ty = VTy;
5264   if (!Ty)
5265     return nullptr;
5266 
5267   auto getAlignmentValue32 = [&](Address addr) -> Value* {
5268     return Builder.getInt32(addr.getAlignment().getQuantity());
5269   };
5270 
5271   unsigned Int = LLVMIntrinsic;
5272   if ((Modifier & UnsignedAlts) && !Usgn)
5273     Int = AltLLVMIntrinsic;
5274 
5275   switch (BuiltinID) {
5276   default: break;
5277   case NEON::BI__builtin_neon_vpadd_v:
5278   case NEON::BI__builtin_neon_vpaddq_v:
5279     // We don't allow fp/int overloading of intrinsics.
5280     if (VTy->getElementType()->isFloatingPointTy() &&
5281         Int == Intrinsic::aarch64_neon_addp)
5282       Int = Intrinsic::aarch64_neon_faddp;
5283     break;
5284   case NEON::BI__builtin_neon_vabs_v:
5285   case NEON::BI__builtin_neon_vabsq_v:
5286     if (VTy->getElementType()->isFloatingPointTy())
5287       return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs");
5288     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs");
5289   case NEON::BI__builtin_neon_vaddhn_v: {
5290     llvm::VectorType *SrcTy =
5291         llvm::VectorType::getExtendedElementVectorType(VTy);
5292 
5293     // %sum = add <4 x i32> %lhs, %rhs
5294     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
5295     Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
5296     Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn");
5297 
5298     // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
5299     Constant *ShiftAmt =
5300         ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2);
5301     Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn");
5302 
5303     // %res = trunc <4 x i32> %high to <4 x i16>
5304     return Builder.CreateTrunc(Ops[0], VTy, "vaddhn");
5305   }
5306   case NEON::BI__builtin_neon_vcale_v:
5307   case NEON::BI__builtin_neon_vcaleq_v:
5308   case NEON::BI__builtin_neon_vcalt_v:
5309   case NEON::BI__builtin_neon_vcaltq_v:
5310     std::swap(Ops[0], Ops[1]);
5311     LLVM_FALLTHROUGH;
5312   case NEON::BI__builtin_neon_vcage_v:
5313   case NEON::BI__builtin_neon_vcageq_v:
5314   case NEON::BI__builtin_neon_vcagt_v:
5315   case NEON::BI__builtin_neon_vcagtq_v: {
5316     llvm::Type *Ty;
5317     switch (VTy->getScalarSizeInBits()) {
5318     default: llvm_unreachable("unexpected type");
5319     case 32:
5320       Ty = FloatTy;
5321       break;
5322     case 64:
5323       Ty = DoubleTy;
5324       break;
5325     case 16:
5326       Ty = HalfTy;
5327       break;
5328     }
5329     llvm::Type *VecFlt = llvm::VectorType::get(Ty, VTy->getNumElements());
5330     llvm::Type *Tys[] = { VTy, VecFlt };
5331     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
5332     return EmitNeonCall(F, Ops, NameHint);
5333   }
5334   case NEON::BI__builtin_neon_vceqz_v:
5335   case NEON::BI__builtin_neon_vceqzq_v:
5336     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ,
5337                                          ICmpInst::ICMP_EQ, "vceqz");
5338   case NEON::BI__builtin_neon_vcgez_v:
5339   case NEON::BI__builtin_neon_vcgezq_v:
5340     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE,
5341                                          ICmpInst::ICMP_SGE, "vcgez");
5342   case NEON::BI__builtin_neon_vclez_v:
5343   case NEON::BI__builtin_neon_vclezq_v:
5344     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE,
5345                                          ICmpInst::ICMP_SLE, "vclez");
5346   case NEON::BI__builtin_neon_vcgtz_v:
5347   case NEON::BI__builtin_neon_vcgtzq_v:
5348     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT,
5349                                          ICmpInst::ICMP_SGT, "vcgtz");
5350   case NEON::BI__builtin_neon_vcltz_v:
5351   case NEON::BI__builtin_neon_vcltzq_v:
5352     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT,
5353                                          ICmpInst::ICMP_SLT, "vcltz");
5354   case NEON::BI__builtin_neon_vclz_v:
5355   case NEON::BI__builtin_neon_vclzq_v:
5356     // We generate target-independent intrinsic, which needs a second argument
5357     // for whether or not clz of zero is undefined; on ARM it isn't.
5358     Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef()));
5359     break;
5360   case NEON::BI__builtin_neon_vcvt_f32_v:
5361   case NEON::BI__builtin_neon_vcvtq_f32_v:
5362     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5363     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad),
5364                      HasLegalHalfType);
5365     return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
5366                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
5367   case NEON::BI__builtin_neon_vcvt_f16_v:
5368   case NEON::BI__builtin_neon_vcvtq_f16_v:
5369     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5370     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float16, false, Quad),
5371                      HasLegalHalfType);
5372     return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
5373                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
5374   case NEON::BI__builtin_neon_vcvt_n_f16_v:
5375   case NEON::BI__builtin_neon_vcvt_n_f32_v:
5376   case NEON::BI__builtin_neon_vcvt_n_f64_v:
5377   case NEON::BI__builtin_neon_vcvtq_n_f16_v:
5378   case NEON::BI__builtin_neon_vcvtq_n_f32_v:
5379   case NEON::BI__builtin_neon_vcvtq_n_f64_v: {
5380     llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty };
5381     Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
5382     Function *F = CGM.getIntrinsic(Int, Tys);
5383     return EmitNeonCall(F, Ops, "vcvt_n");
5384   }
5385   case NEON::BI__builtin_neon_vcvt_n_s16_v:
5386   case NEON::BI__builtin_neon_vcvt_n_s32_v:
5387   case NEON::BI__builtin_neon_vcvt_n_u16_v:
5388   case NEON::BI__builtin_neon_vcvt_n_u32_v:
5389   case NEON::BI__builtin_neon_vcvt_n_s64_v:
5390   case NEON::BI__builtin_neon_vcvt_n_u64_v:
5391   case NEON::BI__builtin_neon_vcvtq_n_s16_v:
5392   case NEON::BI__builtin_neon_vcvtq_n_s32_v:
5393   case NEON::BI__builtin_neon_vcvtq_n_u16_v:
5394   case NEON::BI__builtin_neon_vcvtq_n_u32_v:
5395   case NEON::BI__builtin_neon_vcvtq_n_s64_v:
5396   case NEON::BI__builtin_neon_vcvtq_n_u64_v: {
5397     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
5398     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
5399     return EmitNeonCall(F, Ops, "vcvt_n");
5400   }
5401   case NEON::BI__builtin_neon_vcvt_s32_v:
5402   case NEON::BI__builtin_neon_vcvt_u32_v:
5403   case NEON::BI__builtin_neon_vcvt_s64_v:
5404   case NEON::BI__builtin_neon_vcvt_u64_v:
5405   case NEON::BI__builtin_neon_vcvt_s16_v:
5406   case NEON::BI__builtin_neon_vcvt_u16_v:
5407   case NEON::BI__builtin_neon_vcvtq_s32_v:
5408   case NEON::BI__builtin_neon_vcvtq_u32_v:
5409   case NEON::BI__builtin_neon_vcvtq_s64_v:
5410   case NEON::BI__builtin_neon_vcvtq_u64_v:
5411   case NEON::BI__builtin_neon_vcvtq_s16_v:
5412   case NEON::BI__builtin_neon_vcvtq_u16_v: {
5413     Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type));
5414     return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
5415                 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
5416   }
5417   case NEON::BI__builtin_neon_vcvta_s16_v:
5418   case NEON::BI__builtin_neon_vcvta_s32_v:
5419   case NEON::BI__builtin_neon_vcvta_s64_v:
5420   case NEON::BI__builtin_neon_vcvta_u16_v:
5421   case NEON::BI__builtin_neon_vcvta_u32_v:
5422   case NEON::BI__builtin_neon_vcvta_u64_v:
5423   case NEON::BI__builtin_neon_vcvtaq_s16_v:
5424   case NEON::BI__builtin_neon_vcvtaq_s32_v:
5425   case NEON::BI__builtin_neon_vcvtaq_s64_v:
5426   case NEON::BI__builtin_neon_vcvtaq_u16_v:
5427   case NEON::BI__builtin_neon_vcvtaq_u32_v:
5428   case NEON::BI__builtin_neon_vcvtaq_u64_v:
5429   case NEON::BI__builtin_neon_vcvtn_s16_v:
5430   case NEON::BI__builtin_neon_vcvtn_s32_v:
5431   case NEON::BI__builtin_neon_vcvtn_s64_v:
5432   case NEON::BI__builtin_neon_vcvtn_u16_v:
5433   case NEON::BI__builtin_neon_vcvtn_u32_v:
5434   case NEON::BI__builtin_neon_vcvtn_u64_v:
5435   case NEON::BI__builtin_neon_vcvtnq_s16_v:
5436   case NEON::BI__builtin_neon_vcvtnq_s32_v:
5437   case NEON::BI__builtin_neon_vcvtnq_s64_v:
5438   case NEON::BI__builtin_neon_vcvtnq_u16_v:
5439   case NEON::BI__builtin_neon_vcvtnq_u32_v:
5440   case NEON::BI__builtin_neon_vcvtnq_u64_v:
5441   case NEON::BI__builtin_neon_vcvtp_s16_v:
5442   case NEON::BI__builtin_neon_vcvtp_s32_v:
5443   case NEON::BI__builtin_neon_vcvtp_s64_v:
5444   case NEON::BI__builtin_neon_vcvtp_u16_v:
5445   case NEON::BI__builtin_neon_vcvtp_u32_v:
5446   case NEON::BI__builtin_neon_vcvtp_u64_v:
5447   case NEON::BI__builtin_neon_vcvtpq_s16_v:
5448   case NEON::BI__builtin_neon_vcvtpq_s32_v:
5449   case NEON::BI__builtin_neon_vcvtpq_s64_v:
5450   case NEON::BI__builtin_neon_vcvtpq_u16_v:
5451   case NEON::BI__builtin_neon_vcvtpq_u32_v:
5452   case NEON::BI__builtin_neon_vcvtpq_u64_v:
5453   case NEON::BI__builtin_neon_vcvtm_s16_v:
5454   case NEON::BI__builtin_neon_vcvtm_s32_v:
5455   case NEON::BI__builtin_neon_vcvtm_s64_v:
5456   case NEON::BI__builtin_neon_vcvtm_u16_v:
5457   case NEON::BI__builtin_neon_vcvtm_u32_v:
5458   case NEON::BI__builtin_neon_vcvtm_u64_v:
5459   case NEON::BI__builtin_neon_vcvtmq_s16_v:
5460   case NEON::BI__builtin_neon_vcvtmq_s32_v:
5461   case NEON::BI__builtin_neon_vcvtmq_s64_v:
5462   case NEON::BI__builtin_neon_vcvtmq_u16_v:
5463   case NEON::BI__builtin_neon_vcvtmq_u32_v:
5464   case NEON::BI__builtin_neon_vcvtmq_u64_v: {
5465     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
5466     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
5467   }
5468   case NEON::BI__builtin_neon_vcvtx_f32_v: {
5469     llvm::Type *Tys[2] = { VTy->getTruncatedElementVectorType(VTy), Ty};
5470     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
5471 
5472   }
5473   case NEON::BI__builtin_neon_vext_v:
5474   case NEON::BI__builtin_neon_vextq_v: {
5475     int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
5476     SmallVector<uint32_t, 16> Indices;
5477     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
5478       Indices.push_back(i+CV);
5479 
5480     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5481     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5482     return Builder.CreateShuffleVector(Ops[0], Ops[1], Indices, "vext");
5483   }
5484   case NEON::BI__builtin_neon_vfma_v:
5485   case NEON::BI__builtin_neon_vfmaq_v: {
5486     Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
5487     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5488     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5489     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5490 
5491     // NEON intrinsic puts accumulator first, unlike the LLVM fma.
5492     return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]});
5493   }
5494   case NEON::BI__builtin_neon_vld1_v:
5495   case NEON::BI__builtin_neon_vld1q_v: {
5496     llvm::Type *Tys[] = {Ty, Int8PtrTy};
5497     Ops.push_back(getAlignmentValue32(PtrOp0));
5498     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1");
5499   }
5500   case NEON::BI__builtin_neon_vld1_x2_v:
5501   case NEON::BI__builtin_neon_vld1q_x2_v:
5502   case NEON::BI__builtin_neon_vld1_x3_v:
5503   case NEON::BI__builtin_neon_vld1q_x3_v:
5504   case NEON::BI__builtin_neon_vld1_x4_v:
5505   case NEON::BI__builtin_neon_vld1q_x4_v: {
5506     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType());
5507     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5508     llvm::Type *Tys[2] = { VTy, PTy };
5509     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
5510     Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN");
5511     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5512     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5513     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
5514   }
5515   case NEON::BI__builtin_neon_vld2_v:
5516   case NEON::BI__builtin_neon_vld2q_v:
5517   case NEON::BI__builtin_neon_vld3_v:
5518   case NEON::BI__builtin_neon_vld3q_v:
5519   case NEON::BI__builtin_neon_vld4_v:
5520   case NEON::BI__builtin_neon_vld4q_v:
5521   case NEON::BI__builtin_neon_vld2_dup_v:
5522   case NEON::BI__builtin_neon_vld2q_dup_v:
5523   case NEON::BI__builtin_neon_vld3_dup_v:
5524   case NEON::BI__builtin_neon_vld3q_dup_v:
5525   case NEON::BI__builtin_neon_vld4_dup_v:
5526   case NEON::BI__builtin_neon_vld4q_dup_v: {
5527     llvm::Type *Tys[] = {Ty, Int8PtrTy};
5528     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
5529     Value *Align = getAlignmentValue32(PtrOp1);
5530     Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint);
5531     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5532     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5533     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
5534   }
5535   case NEON::BI__builtin_neon_vld1_dup_v:
5536   case NEON::BI__builtin_neon_vld1q_dup_v: {
5537     Value *V = UndefValue::get(Ty);
5538     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
5539     PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty);
5540     LoadInst *Ld = Builder.CreateLoad(PtrOp0);
5541     llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
5542     Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
5543     return EmitNeonSplat(Ops[0], CI);
5544   }
5545   case NEON::BI__builtin_neon_vld2_lane_v:
5546   case NEON::BI__builtin_neon_vld2q_lane_v:
5547   case NEON::BI__builtin_neon_vld3_lane_v:
5548   case NEON::BI__builtin_neon_vld3q_lane_v:
5549   case NEON::BI__builtin_neon_vld4_lane_v:
5550   case NEON::BI__builtin_neon_vld4q_lane_v: {
5551     llvm::Type *Tys[] = {Ty, Int8PtrTy};
5552     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
5553     for (unsigned I = 2; I < Ops.size() - 1; ++I)
5554       Ops[I] = Builder.CreateBitCast(Ops[I], Ty);
5555     Ops.push_back(getAlignmentValue32(PtrOp1));
5556     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint);
5557     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5558     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5559     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
5560   }
5561   case NEON::BI__builtin_neon_vmovl_v: {
5562     llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy);
5563     Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
5564     if (Usgn)
5565       return Builder.CreateZExt(Ops[0], Ty, "vmovl");
5566     return Builder.CreateSExt(Ops[0], Ty, "vmovl");
5567   }
5568   case NEON::BI__builtin_neon_vmovn_v: {
5569     llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy);
5570     Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
5571     return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
5572   }
5573   case NEON::BI__builtin_neon_vmull_v:
5574     // FIXME: the integer vmull operations could be emitted in terms of pure
5575     // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of
5576     // hoisting the exts outside loops. Until global ISel comes along that can
5577     // see through such movement this leads to bad CodeGen. So we need an
5578     // intrinsic for now.
5579     Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
5580     Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
5581     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
5582   case NEON::BI__builtin_neon_vpadal_v:
5583   case NEON::BI__builtin_neon_vpadalq_v: {
5584     // The source operand type has twice as many elements of half the size.
5585     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
5586     llvm::Type *EltTy =
5587       llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
5588     llvm::Type *NarrowTy =
5589       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
5590     llvm::Type *Tys[2] = { Ty, NarrowTy };
5591     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
5592   }
5593   case NEON::BI__builtin_neon_vpaddl_v:
5594   case NEON::BI__builtin_neon_vpaddlq_v: {
5595     // The source operand type has twice as many elements of half the size.
5596     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
5597     llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
5598     llvm::Type *NarrowTy =
5599       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
5600     llvm::Type *Tys[2] = { Ty, NarrowTy };
5601     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
5602   }
5603   case NEON::BI__builtin_neon_vqdmlal_v:
5604   case NEON::BI__builtin_neon_vqdmlsl_v: {
5605     SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end());
5606     Ops[1] =
5607         EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal");
5608     Ops.resize(2);
5609     return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint);
5610   }
5611   case NEON::BI__builtin_neon_vqshl_n_v:
5612   case NEON::BI__builtin_neon_vqshlq_n_v:
5613     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
5614                         1, false);
5615   case NEON::BI__builtin_neon_vqshlu_n_v:
5616   case NEON::BI__builtin_neon_vqshluq_n_v:
5617     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n",
5618                         1, false);
5619   case NEON::BI__builtin_neon_vrecpe_v:
5620   case NEON::BI__builtin_neon_vrecpeq_v:
5621   case NEON::BI__builtin_neon_vrsqrte_v:
5622   case NEON::BI__builtin_neon_vrsqrteq_v:
5623     Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic;
5624     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
5625   case NEON::BI__builtin_neon_vrndi_v:
5626   case NEON::BI__builtin_neon_vrndiq_v:
5627     Int = Intrinsic::nearbyint;
5628     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
5629   case NEON::BI__builtin_neon_vrshr_n_v:
5630   case NEON::BI__builtin_neon_vrshrq_n_v:
5631     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n",
5632                         1, true);
5633   case NEON::BI__builtin_neon_vshl_n_v:
5634   case NEON::BI__builtin_neon_vshlq_n_v:
5635     Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
5636     return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
5637                              "vshl_n");
5638   case NEON::BI__builtin_neon_vshll_n_v: {
5639     llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy);
5640     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
5641     if (Usgn)
5642       Ops[0] = Builder.CreateZExt(Ops[0], VTy);
5643     else
5644       Ops[0] = Builder.CreateSExt(Ops[0], VTy);
5645     Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false);
5646     return Builder.CreateShl(Ops[0], Ops[1], "vshll_n");
5647   }
5648   case NEON::BI__builtin_neon_vshrn_n_v: {
5649     llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy);
5650     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
5651     Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false);
5652     if (Usgn)
5653       Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]);
5654     else
5655       Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]);
5656     return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n");
5657   }
5658   case NEON::BI__builtin_neon_vshr_n_v:
5659   case NEON::BI__builtin_neon_vshrq_n_v:
5660     return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n");
5661   case NEON::BI__builtin_neon_vst1_v:
5662   case NEON::BI__builtin_neon_vst1q_v:
5663   case NEON::BI__builtin_neon_vst2_v:
5664   case NEON::BI__builtin_neon_vst2q_v:
5665   case NEON::BI__builtin_neon_vst3_v:
5666   case NEON::BI__builtin_neon_vst3q_v:
5667   case NEON::BI__builtin_neon_vst4_v:
5668   case NEON::BI__builtin_neon_vst4q_v:
5669   case NEON::BI__builtin_neon_vst2_lane_v:
5670   case NEON::BI__builtin_neon_vst2q_lane_v:
5671   case NEON::BI__builtin_neon_vst3_lane_v:
5672   case NEON::BI__builtin_neon_vst3q_lane_v:
5673   case NEON::BI__builtin_neon_vst4_lane_v:
5674   case NEON::BI__builtin_neon_vst4q_lane_v: {
5675     llvm::Type *Tys[] = {Int8PtrTy, Ty};
5676     Ops.push_back(getAlignmentValue32(PtrOp0));
5677     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "");
5678   }
5679   case NEON::BI__builtin_neon_vst1_x2_v:
5680   case NEON::BI__builtin_neon_vst1q_x2_v:
5681   case NEON::BI__builtin_neon_vst1_x3_v:
5682   case NEON::BI__builtin_neon_vst1q_x3_v:
5683   case NEON::BI__builtin_neon_vst1_x4_v:
5684   case NEON::BI__builtin_neon_vst1q_x4_v: {
5685     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType());
5686     // TODO: Currently in AArch32 mode the pointer operand comes first, whereas
5687     // in AArch64 it comes last. We may want to stick to one or another.
5688     if (Arch == llvm::Triple::aarch64 || Arch == llvm::Triple::aarch64_be ||
5689         Arch == llvm::Triple::aarch64_32) {
5690       llvm::Type *Tys[2] = { VTy, PTy };
5691       std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
5692       return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "");
5693     }
5694     llvm::Type *Tys[2] = { PTy, VTy };
5695     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "");
5696   }
5697   case NEON::BI__builtin_neon_vsubhn_v: {
5698     llvm::VectorType *SrcTy =
5699         llvm::VectorType::getExtendedElementVectorType(VTy);
5700 
5701     // %sum = add <4 x i32> %lhs, %rhs
5702     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
5703     Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
5704     Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn");
5705 
5706     // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
5707     Constant *ShiftAmt =
5708         ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2);
5709     Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn");
5710 
5711     // %res = trunc <4 x i32> %high to <4 x i16>
5712     return Builder.CreateTrunc(Ops[0], VTy, "vsubhn");
5713   }
5714   case NEON::BI__builtin_neon_vtrn_v:
5715   case NEON::BI__builtin_neon_vtrnq_v: {
5716     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5717     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5718     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5719     Value *SV = nullptr;
5720 
5721     for (unsigned vi = 0; vi != 2; ++vi) {
5722       SmallVector<uint32_t, 16> Indices;
5723       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
5724         Indices.push_back(i+vi);
5725         Indices.push_back(i+e+vi);
5726       }
5727       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
5728       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn");
5729       SV = Builder.CreateDefaultAlignedStore(SV, Addr);
5730     }
5731     return SV;
5732   }
5733   case NEON::BI__builtin_neon_vtst_v:
5734   case NEON::BI__builtin_neon_vtstq_v: {
5735     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5736     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5737     Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
5738     Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
5739                                 ConstantAggregateZero::get(Ty));
5740     return Builder.CreateSExt(Ops[0], Ty, "vtst");
5741   }
5742   case NEON::BI__builtin_neon_vuzp_v:
5743   case NEON::BI__builtin_neon_vuzpq_v: {
5744     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5745     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5746     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5747     Value *SV = nullptr;
5748 
5749     for (unsigned vi = 0; vi != 2; ++vi) {
5750       SmallVector<uint32_t, 16> Indices;
5751       for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
5752         Indices.push_back(2*i+vi);
5753 
5754       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
5755       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp");
5756       SV = Builder.CreateDefaultAlignedStore(SV, Addr);
5757     }
5758     return SV;
5759   }
5760   case NEON::BI__builtin_neon_vzip_v:
5761   case NEON::BI__builtin_neon_vzipq_v: {
5762     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5763     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5764     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5765     Value *SV = nullptr;
5766 
5767     for (unsigned vi = 0; vi != 2; ++vi) {
5768       SmallVector<uint32_t, 16> Indices;
5769       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
5770         Indices.push_back((i + vi*e) >> 1);
5771         Indices.push_back(((i + vi*e) >> 1)+e);
5772       }
5773       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
5774       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip");
5775       SV = Builder.CreateDefaultAlignedStore(SV, Addr);
5776     }
5777     return SV;
5778   }
5779   case NEON::BI__builtin_neon_vdot_v:
5780   case NEON::BI__builtin_neon_vdotq_v: {
5781     llvm::Type *InputTy =
5782         llvm::VectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
5783     llvm::Type *Tys[2] = { Ty, InputTy };
5784     Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
5785     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vdot");
5786   }
5787   case NEON::BI__builtin_neon_vfmlal_low_v:
5788   case NEON::BI__builtin_neon_vfmlalq_low_v: {
5789     llvm::Type *InputTy =
5790         llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
5791     llvm::Type *Tys[2] = { Ty, InputTy };
5792     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_low");
5793   }
5794   case NEON::BI__builtin_neon_vfmlsl_low_v:
5795   case NEON::BI__builtin_neon_vfmlslq_low_v: {
5796     llvm::Type *InputTy =
5797         llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
5798     llvm::Type *Tys[2] = { Ty, InputTy };
5799     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_low");
5800   }
5801   case NEON::BI__builtin_neon_vfmlal_high_v:
5802   case NEON::BI__builtin_neon_vfmlalq_high_v: {
5803     llvm::Type *InputTy =
5804            llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
5805     llvm::Type *Tys[2] = { Ty, InputTy };
5806     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_high");
5807   }
5808   case NEON::BI__builtin_neon_vfmlsl_high_v:
5809   case NEON::BI__builtin_neon_vfmlslq_high_v: {
5810     llvm::Type *InputTy =
5811            llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
5812     llvm::Type *Tys[2] = { Ty, InputTy };
5813     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_high");
5814   }
5815   }
5816 
5817   assert(Int && "Expected valid intrinsic number");
5818 
5819   // Determine the type(s) of this overloaded AArch64 intrinsic.
5820   Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E);
5821 
5822   Value *Result = EmitNeonCall(F, Ops, NameHint);
5823   llvm::Type *ResultType = ConvertType(E->getType());
5824   // AArch64 intrinsic one-element vector type cast to
5825   // scalar type expected by the builtin
5826   return Builder.CreateBitCast(Result, ResultType, NameHint);
5827 }
5828 
5829 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr(
5830     Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp,
5831     const CmpInst::Predicate Ip, const Twine &Name) {
5832   llvm::Type *OTy = Op->getType();
5833 
5834   // FIXME: this is utterly horrific. We should not be looking at previous
5835   // codegen context to find out what needs doing. Unfortunately TableGen
5836   // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32
5837   // (etc).
5838   if (BitCastInst *BI = dyn_cast<BitCastInst>(Op))
5839     OTy = BI->getOperand(0)->getType();
5840 
5841   Op = Builder.CreateBitCast(Op, OTy);
5842   if (OTy->getScalarType()->isFloatingPointTy()) {
5843     Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy));
5844   } else {
5845     Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy));
5846   }
5847   return Builder.CreateSExt(Op, Ty, Name);
5848 }
5849 
5850 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
5851                                  Value *ExtOp, Value *IndexOp,
5852                                  llvm::Type *ResTy, unsigned IntID,
5853                                  const char *Name) {
5854   SmallVector<Value *, 2> TblOps;
5855   if (ExtOp)
5856     TblOps.push_back(ExtOp);
5857 
5858   // Build a vector containing sequential number like (0, 1, 2, ..., 15)
5859   SmallVector<uint32_t, 16> Indices;
5860   llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType());
5861   for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) {
5862     Indices.push_back(2*i);
5863     Indices.push_back(2*i+1);
5864   }
5865 
5866   int PairPos = 0, End = Ops.size() - 1;
5867   while (PairPos < End) {
5868     TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
5869                                                      Ops[PairPos+1], Indices,
5870                                                      Name));
5871     PairPos += 2;
5872   }
5873 
5874   // If there's an odd number of 64-bit lookup table, fill the high 64-bit
5875   // of the 128-bit lookup table with zero.
5876   if (PairPos == End) {
5877     Value *ZeroTbl = ConstantAggregateZero::get(TblTy);
5878     TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
5879                                                      ZeroTbl, Indices, Name));
5880   }
5881 
5882   Function *TblF;
5883   TblOps.push_back(IndexOp);
5884   TblF = CGF.CGM.getIntrinsic(IntID, ResTy);
5885 
5886   return CGF.EmitNeonCall(TblF, TblOps, Name);
5887 }
5888 
5889 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) {
5890   unsigned Value;
5891   switch (BuiltinID) {
5892   default:
5893     return nullptr;
5894   case ARM::BI__builtin_arm_nop:
5895     Value = 0;
5896     break;
5897   case ARM::BI__builtin_arm_yield:
5898   case ARM::BI__yield:
5899     Value = 1;
5900     break;
5901   case ARM::BI__builtin_arm_wfe:
5902   case ARM::BI__wfe:
5903     Value = 2;
5904     break;
5905   case ARM::BI__builtin_arm_wfi:
5906   case ARM::BI__wfi:
5907     Value = 3;
5908     break;
5909   case ARM::BI__builtin_arm_sev:
5910   case ARM::BI__sev:
5911     Value = 4;
5912     break;
5913   case ARM::BI__builtin_arm_sevl:
5914   case ARM::BI__sevl:
5915     Value = 5;
5916     break;
5917   }
5918 
5919   return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
5920                             llvm::ConstantInt::get(Int32Ty, Value));
5921 }
5922 
5923 // Generates the IR for the read/write special register builtin,
5924 // ValueType is the type of the value that is to be written or read,
5925 // RegisterType is the type of the register being written to or read from.
5926 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF,
5927                                          const CallExpr *E,
5928                                          llvm::Type *RegisterType,
5929                                          llvm::Type *ValueType,
5930                                          bool IsRead,
5931                                          StringRef SysReg = "") {
5932   // write and register intrinsics only support 32 and 64 bit operations.
5933   assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64))
5934           && "Unsupported size for register.");
5935 
5936   CodeGen::CGBuilderTy &Builder = CGF.Builder;
5937   CodeGen::CodeGenModule &CGM = CGF.CGM;
5938   LLVMContext &Context = CGM.getLLVMContext();
5939 
5940   if (SysReg.empty()) {
5941     const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts();
5942     SysReg = cast<clang::StringLiteral>(SysRegStrExpr)->getString();
5943   }
5944 
5945   llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) };
5946   llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
5947   llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
5948 
5949   llvm::Type *Types[] = { RegisterType };
5950 
5951   bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32);
5952   assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64))
5953             && "Can't fit 64-bit value in 32-bit register");
5954 
5955   if (IsRead) {
5956     llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types);
5957     llvm::Value *Call = Builder.CreateCall(F, Metadata);
5958 
5959     if (MixedTypes)
5960       // Read into 64 bit register and then truncate result to 32 bit.
5961       return Builder.CreateTrunc(Call, ValueType);
5962 
5963     if (ValueType->isPointerTy())
5964       // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*).
5965       return Builder.CreateIntToPtr(Call, ValueType);
5966 
5967     return Call;
5968   }
5969 
5970   llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types);
5971   llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1));
5972   if (MixedTypes) {
5973     // Extend 32 bit write value to 64 bit to pass to write.
5974     ArgValue = Builder.CreateZExt(ArgValue, RegisterType);
5975     return Builder.CreateCall(F, { Metadata, ArgValue });
5976   }
5977 
5978   if (ValueType->isPointerTy()) {
5979     // Have VoidPtrTy ArgValue but want to return an i32/i64.
5980     ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType);
5981     return Builder.CreateCall(F, { Metadata, ArgValue });
5982   }
5983 
5984   return Builder.CreateCall(F, { Metadata, ArgValue });
5985 }
5986 
5987 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra
5988 /// argument that specifies the vector type.
5989 static bool HasExtraNeonArgument(unsigned BuiltinID) {
5990   switch (BuiltinID) {
5991   default: break;
5992   case NEON::BI__builtin_neon_vget_lane_i8:
5993   case NEON::BI__builtin_neon_vget_lane_i16:
5994   case NEON::BI__builtin_neon_vget_lane_i32:
5995   case NEON::BI__builtin_neon_vget_lane_i64:
5996   case NEON::BI__builtin_neon_vget_lane_f32:
5997   case NEON::BI__builtin_neon_vgetq_lane_i8:
5998   case NEON::BI__builtin_neon_vgetq_lane_i16:
5999   case NEON::BI__builtin_neon_vgetq_lane_i32:
6000   case NEON::BI__builtin_neon_vgetq_lane_i64:
6001   case NEON::BI__builtin_neon_vgetq_lane_f32:
6002   case NEON::BI__builtin_neon_vset_lane_i8:
6003   case NEON::BI__builtin_neon_vset_lane_i16:
6004   case NEON::BI__builtin_neon_vset_lane_i32:
6005   case NEON::BI__builtin_neon_vset_lane_i64:
6006   case NEON::BI__builtin_neon_vset_lane_f32:
6007   case NEON::BI__builtin_neon_vsetq_lane_i8:
6008   case NEON::BI__builtin_neon_vsetq_lane_i16:
6009   case NEON::BI__builtin_neon_vsetq_lane_i32:
6010   case NEON::BI__builtin_neon_vsetq_lane_i64:
6011   case NEON::BI__builtin_neon_vsetq_lane_f32:
6012   case NEON::BI__builtin_neon_vsha1h_u32:
6013   case NEON::BI__builtin_neon_vsha1cq_u32:
6014   case NEON::BI__builtin_neon_vsha1pq_u32:
6015   case NEON::BI__builtin_neon_vsha1mq_u32:
6016   case clang::ARM::BI_MoveToCoprocessor:
6017   case clang::ARM::BI_MoveToCoprocessor2:
6018     return false;
6019   }
6020   return true;
6021 }
6022 
6023 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
6024                                            const CallExpr *E,
6025                                            ReturnValueSlot ReturnValue,
6026                                            llvm::Triple::ArchType Arch) {
6027   if (auto Hint = GetValueForARMHint(BuiltinID))
6028     return Hint;
6029 
6030   if (BuiltinID == ARM::BI__emit) {
6031     bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb;
6032     llvm::FunctionType *FTy =
6033         llvm::FunctionType::get(VoidTy, /*Variadic=*/false);
6034 
6035     Expr::EvalResult Result;
6036     if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
6037       llvm_unreachable("Sema will ensure that the parameter is constant");
6038 
6039     llvm::APSInt Value = Result.Val.getInt();
6040     uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue();
6041 
6042     llvm::InlineAsm *Emit =
6043         IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "",
6044                                  /*hasSideEffects=*/true)
6045                 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "",
6046                                  /*hasSideEffects=*/true);
6047 
6048     return Builder.CreateCall(Emit);
6049   }
6050 
6051   if (BuiltinID == ARM::BI__builtin_arm_dbg) {
6052     Value *Option = EmitScalarExpr(E->getArg(0));
6053     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option);
6054   }
6055 
6056   if (BuiltinID == ARM::BI__builtin_arm_prefetch) {
6057     Value *Address = EmitScalarExpr(E->getArg(0));
6058     Value *RW      = EmitScalarExpr(E->getArg(1));
6059     Value *IsData  = EmitScalarExpr(E->getArg(2));
6060 
6061     // Locality is not supported on ARM target
6062     Value *Locality = llvm::ConstantInt::get(Int32Ty, 3);
6063 
6064     Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
6065     return Builder.CreateCall(F, {Address, RW, Locality, IsData});
6066   }
6067 
6068   if (BuiltinID == ARM::BI__builtin_arm_rbit) {
6069     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
6070     return Builder.CreateCall(
6071         CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
6072   }
6073 
6074   if (BuiltinID == ARM::BI__builtin_arm_cls) {
6075     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
6076     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls), Arg, "cls");
6077   }
6078   if (BuiltinID == ARM::BI__builtin_arm_cls64) {
6079     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
6080     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls64), Arg,
6081                               "cls");
6082   }
6083 
6084   if (BuiltinID == ARM::BI__clear_cache) {
6085     assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
6086     const FunctionDecl *FD = E->getDirectCallee();
6087     Value *Ops[2];
6088     for (unsigned i = 0; i < 2; i++)
6089       Ops[i] = EmitScalarExpr(E->getArg(i));
6090     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
6091     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
6092     StringRef Name = FD->getName();
6093     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
6094   }
6095 
6096   if (BuiltinID == ARM::BI__builtin_arm_mcrr ||
6097       BuiltinID == ARM::BI__builtin_arm_mcrr2) {
6098     Function *F;
6099 
6100     switch (BuiltinID) {
6101     default: llvm_unreachable("unexpected builtin");
6102     case ARM::BI__builtin_arm_mcrr:
6103       F = CGM.getIntrinsic(Intrinsic::arm_mcrr);
6104       break;
6105     case ARM::BI__builtin_arm_mcrr2:
6106       F = CGM.getIntrinsic(Intrinsic::arm_mcrr2);
6107       break;
6108     }
6109 
6110     // MCRR{2} instruction has 5 operands but
6111     // the intrinsic has 4 because Rt and Rt2
6112     // are represented as a single unsigned 64
6113     // bit integer in the intrinsic definition
6114     // but internally it's represented as 2 32
6115     // bit integers.
6116 
6117     Value *Coproc = EmitScalarExpr(E->getArg(0));
6118     Value *Opc1 = EmitScalarExpr(E->getArg(1));
6119     Value *RtAndRt2 = EmitScalarExpr(E->getArg(2));
6120     Value *CRm = EmitScalarExpr(E->getArg(3));
6121 
6122     Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
6123     Value *Rt = Builder.CreateTruncOrBitCast(RtAndRt2, Int32Ty);
6124     Value *Rt2 = Builder.CreateLShr(RtAndRt2, C1);
6125     Rt2 = Builder.CreateTruncOrBitCast(Rt2, Int32Ty);
6126 
6127     return Builder.CreateCall(F, {Coproc, Opc1, Rt, Rt2, CRm});
6128   }
6129 
6130   if (BuiltinID == ARM::BI__builtin_arm_mrrc ||
6131       BuiltinID == ARM::BI__builtin_arm_mrrc2) {
6132     Function *F;
6133 
6134     switch (BuiltinID) {
6135     default: llvm_unreachable("unexpected builtin");
6136     case ARM::BI__builtin_arm_mrrc:
6137       F = CGM.getIntrinsic(Intrinsic::arm_mrrc);
6138       break;
6139     case ARM::BI__builtin_arm_mrrc2:
6140       F = CGM.getIntrinsic(Intrinsic::arm_mrrc2);
6141       break;
6142     }
6143 
6144     Value *Coproc = EmitScalarExpr(E->getArg(0));
6145     Value *Opc1 = EmitScalarExpr(E->getArg(1));
6146     Value *CRm  = EmitScalarExpr(E->getArg(2));
6147     Value *RtAndRt2 = Builder.CreateCall(F, {Coproc, Opc1, CRm});
6148 
6149     // Returns an unsigned 64 bit integer, represented
6150     // as two 32 bit integers.
6151 
6152     Value *Rt = Builder.CreateExtractValue(RtAndRt2, 1);
6153     Value *Rt1 = Builder.CreateExtractValue(RtAndRt2, 0);
6154     Rt = Builder.CreateZExt(Rt, Int64Ty);
6155     Rt1 = Builder.CreateZExt(Rt1, Int64Ty);
6156 
6157     Value *ShiftCast = llvm::ConstantInt::get(Int64Ty, 32);
6158     RtAndRt2 = Builder.CreateShl(Rt, ShiftCast, "shl", true);
6159     RtAndRt2 = Builder.CreateOr(RtAndRt2, Rt1);
6160 
6161     return Builder.CreateBitCast(RtAndRt2, ConvertType(E->getType()));
6162   }
6163 
6164   if (BuiltinID == ARM::BI__builtin_arm_ldrexd ||
6165       ((BuiltinID == ARM::BI__builtin_arm_ldrex ||
6166         BuiltinID == ARM::BI__builtin_arm_ldaex) &&
6167        getContext().getTypeSize(E->getType()) == 64) ||
6168       BuiltinID == ARM::BI__ldrexd) {
6169     Function *F;
6170 
6171     switch (BuiltinID) {
6172     default: llvm_unreachable("unexpected builtin");
6173     case ARM::BI__builtin_arm_ldaex:
6174       F = CGM.getIntrinsic(Intrinsic::arm_ldaexd);
6175       break;
6176     case ARM::BI__builtin_arm_ldrexd:
6177     case ARM::BI__builtin_arm_ldrex:
6178     case ARM::BI__ldrexd:
6179       F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
6180       break;
6181     }
6182 
6183     Value *LdPtr = EmitScalarExpr(E->getArg(0));
6184     Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
6185                                     "ldrexd");
6186 
6187     Value *Val0 = Builder.CreateExtractValue(Val, 1);
6188     Value *Val1 = Builder.CreateExtractValue(Val, 0);
6189     Val0 = Builder.CreateZExt(Val0, Int64Ty);
6190     Val1 = Builder.CreateZExt(Val1, Int64Ty);
6191 
6192     Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
6193     Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
6194     Val = Builder.CreateOr(Val, Val1);
6195     return Builder.CreateBitCast(Val, ConvertType(E->getType()));
6196   }
6197 
6198   if (BuiltinID == ARM::BI__builtin_arm_ldrex ||
6199       BuiltinID == ARM::BI__builtin_arm_ldaex) {
6200     Value *LoadAddr = EmitScalarExpr(E->getArg(0));
6201 
6202     QualType Ty = E->getType();
6203     llvm::Type *RealResTy = ConvertType(Ty);
6204     llvm::Type *PtrTy = llvm::IntegerType::get(
6205         getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo();
6206     LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy);
6207 
6208     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex
6209                                        ? Intrinsic::arm_ldaex
6210                                        : Intrinsic::arm_ldrex,
6211                                    PtrTy);
6212     Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex");
6213 
6214     if (RealResTy->isPointerTy())
6215       return Builder.CreateIntToPtr(Val, RealResTy);
6216     else {
6217       llvm::Type *IntResTy = llvm::IntegerType::get(
6218           getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy));
6219       Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
6220       return Builder.CreateBitCast(Val, RealResTy);
6221     }
6222   }
6223 
6224   if (BuiltinID == ARM::BI__builtin_arm_strexd ||
6225       ((BuiltinID == ARM::BI__builtin_arm_stlex ||
6226         BuiltinID == ARM::BI__builtin_arm_strex) &&
6227        getContext().getTypeSize(E->getArg(0)->getType()) == 64)) {
6228     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
6229                                        ? Intrinsic::arm_stlexd
6230                                        : Intrinsic::arm_strexd);
6231     llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty);
6232 
6233     Address Tmp = CreateMemTemp(E->getArg(0)->getType());
6234     Value *Val = EmitScalarExpr(E->getArg(0));
6235     Builder.CreateStore(Val, Tmp);
6236 
6237     Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
6238     Val = Builder.CreateLoad(LdPtr);
6239 
6240     Value *Arg0 = Builder.CreateExtractValue(Val, 0);
6241     Value *Arg1 = Builder.CreateExtractValue(Val, 1);
6242     Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy);
6243     return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd");
6244   }
6245 
6246   if (BuiltinID == ARM::BI__builtin_arm_strex ||
6247       BuiltinID == ARM::BI__builtin_arm_stlex) {
6248     Value *StoreVal = EmitScalarExpr(E->getArg(0));
6249     Value *StoreAddr = EmitScalarExpr(E->getArg(1));
6250 
6251     QualType Ty = E->getArg(0)->getType();
6252     llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
6253                                                  getContext().getTypeSize(Ty));
6254     StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
6255 
6256     if (StoreVal->getType()->isPointerTy())
6257       StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty);
6258     else {
6259       llvm::Type *IntTy = llvm::IntegerType::get(
6260           getLLVMContext(),
6261           CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType()));
6262       StoreVal = Builder.CreateBitCast(StoreVal, IntTy);
6263       StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty);
6264     }
6265 
6266     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
6267                                        ? Intrinsic::arm_stlex
6268                                        : Intrinsic::arm_strex,
6269                                    StoreAddr->getType());
6270     return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex");
6271   }
6272 
6273   if (BuiltinID == ARM::BI__builtin_arm_clrex) {
6274     Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex);
6275     return Builder.CreateCall(F);
6276   }
6277 
6278   // CRC32
6279   Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
6280   switch (BuiltinID) {
6281   case ARM::BI__builtin_arm_crc32b:
6282     CRCIntrinsicID = Intrinsic::arm_crc32b; break;
6283   case ARM::BI__builtin_arm_crc32cb:
6284     CRCIntrinsicID = Intrinsic::arm_crc32cb; break;
6285   case ARM::BI__builtin_arm_crc32h:
6286     CRCIntrinsicID = Intrinsic::arm_crc32h; break;
6287   case ARM::BI__builtin_arm_crc32ch:
6288     CRCIntrinsicID = Intrinsic::arm_crc32ch; break;
6289   case ARM::BI__builtin_arm_crc32w:
6290   case ARM::BI__builtin_arm_crc32d:
6291     CRCIntrinsicID = Intrinsic::arm_crc32w; break;
6292   case ARM::BI__builtin_arm_crc32cw:
6293   case ARM::BI__builtin_arm_crc32cd:
6294     CRCIntrinsicID = Intrinsic::arm_crc32cw; break;
6295   }
6296 
6297   if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
6298     Value *Arg0 = EmitScalarExpr(E->getArg(0));
6299     Value *Arg1 = EmitScalarExpr(E->getArg(1));
6300 
6301     // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w
6302     // intrinsics, hence we need different codegen for these cases.
6303     if (BuiltinID == ARM::BI__builtin_arm_crc32d ||
6304         BuiltinID == ARM::BI__builtin_arm_crc32cd) {
6305       Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
6306       Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty);
6307       Value *Arg1b = Builder.CreateLShr(Arg1, C1);
6308       Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty);
6309 
6310       Function *F = CGM.getIntrinsic(CRCIntrinsicID);
6311       Value *Res = Builder.CreateCall(F, {Arg0, Arg1a});
6312       return Builder.CreateCall(F, {Res, Arg1b});
6313     } else {
6314       Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty);
6315 
6316       Function *F = CGM.getIntrinsic(CRCIntrinsicID);
6317       return Builder.CreateCall(F, {Arg0, Arg1});
6318     }
6319   }
6320 
6321   if (BuiltinID == ARM::BI__builtin_arm_rsr ||
6322       BuiltinID == ARM::BI__builtin_arm_rsr64 ||
6323       BuiltinID == ARM::BI__builtin_arm_rsrp ||
6324       BuiltinID == ARM::BI__builtin_arm_wsr ||
6325       BuiltinID == ARM::BI__builtin_arm_wsr64 ||
6326       BuiltinID == ARM::BI__builtin_arm_wsrp) {
6327 
6328     bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr ||
6329                   BuiltinID == ARM::BI__builtin_arm_rsr64 ||
6330                   BuiltinID == ARM::BI__builtin_arm_rsrp;
6331 
6332     bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp ||
6333                             BuiltinID == ARM::BI__builtin_arm_wsrp;
6334 
6335     bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 ||
6336                    BuiltinID == ARM::BI__builtin_arm_wsr64;
6337 
6338     llvm::Type *ValueType;
6339     llvm::Type *RegisterType;
6340     if (IsPointerBuiltin) {
6341       ValueType = VoidPtrTy;
6342       RegisterType = Int32Ty;
6343     } else if (Is64Bit) {
6344       ValueType = RegisterType = Int64Ty;
6345     } else {
6346       ValueType = RegisterType = Int32Ty;
6347     }
6348 
6349     return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead);
6350   }
6351 
6352   // Deal with MVE builtins
6353   if (Value *Result = EmitARMMVEBuiltinExpr(BuiltinID, E, ReturnValue, Arch))
6354     return Result;
6355 
6356   // Find out if any arguments are required to be integer constant
6357   // expressions.
6358   unsigned ICEArguments = 0;
6359   ASTContext::GetBuiltinTypeError Error;
6360   getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
6361   assert(Error == ASTContext::GE_None && "Should not codegen an error");
6362 
6363   auto getAlignmentValue32 = [&](Address addr) -> Value* {
6364     return Builder.getInt32(addr.getAlignment().getQuantity());
6365   };
6366 
6367   Address PtrOp0 = Address::invalid();
6368   Address PtrOp1 = Address::invalid();
6369   SmallVector<Value*, 4> Ops;
6370   bool HasExtraArg = HasExtraNeonArgument(BuiltinID);
6371   unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0);
6372   for (unsigned i = 0, e = NumArgs; i != e; i++) {
6373     if (i == 0) {
6374       switch (BuiltinID) {
6375       case NEON::BI__builtin_neon_vld1_v:
6376       case NEON::BI__builtin_neon_vld1q_v:
6377       case NEON::BI__builtin_neon_vld1q_lane_v:
6378       case NEON::BI__builtin_neon_vld1_lane_v:
6379       case NEON::BI__builtin_neon_vld1_dup_v:
6380       case NEON::BI__builtin_neon_vld1q_dup_v:
6381       case NEON::BI__builtin_neon_vst1_v:
6382       case NEON::BI__builtin_neon_vst1q_v:
6383       case NEON::BI__builtin_neon_vst1q_lane_v:
6384       case NEON::BI__builtin_neon_vst1_lane_v:
6385       case NEON::BI__builtin_neon_vst2_v:
6386       case NEON::BI__builtin_neon_vst2q_v:
6387       case NEON::BI__builtin_neon_vst2_lane_v:
6388       case NEON::BI__builtin_neon_vst2q_lane_v:
6389       case NEON::BI__builtin_neon_vst3_v:
6390       case NEON::BI__builtin_neon_vst3q_v:
6391       case NEON::BI__builtin_neon_vst3_lane_v:
6392       case NEON::BI__builtin_neon_vst3q_lane_v:
6393       case NEON::BI__builtin_neon_vst4_v:
6394       case NEON::BI__builtin_neon_vst4q_v:
6395       case NEON::BI__builtin_neon_vst4_lane_v:
6396       case NEON::BI__builtin_neon_vst4q_lane_v:
6397         // Get the alignment for the argument in addition to the value;
6398         // we'll use it later.
6399         PtrOp0 = EmitPointerWithAlignment(E->getArg(0));
6400         Ops.push_back(PtrOp0.getPointer());
6401         continue;
6402       }
6403     }
6404     if (i == 1) {
6405       switch (BuiltinID) {
6406       case NEON::BI__builtin_neon_vld2_v:
6407       case NEON::BI__builtin_neon_vld2q_v:
6408       case NEON::BI__builtin_neon_vld3_v:
6409       case NEON::BI__builtin_neon_vld3q_v:
6410       case NEON::BI__builtin_neon_vld4_v:
6411       case NEON::BI__builtin_neon_vld4q_v:
6412       case NEON::BI__builtin_neon_vld2_lane_v:
6413       case NEON::BI__builtin_neon_vld2q_lane_v:
6414       case NEON::BI__builtin_neon_vld3_lane_v:
6415       case NEON::BI__builtin_neon_vld3q_lane_v:
6416       case NEON::BI__builtin_neon_vld4_lane_v:
6417       case NEON::BI__builtin_neon_vld4q_lane_v:
6418       case NEON::BI__builtin_neon_vld2_dup_v:
6419       case NEON::BI__builtin_neon_vld2q_dup_v:
6420       case NEON::BI__builtin_neon_vld3_dup_v:
6421       case NEON::BI__builtin_neon_vld3q_dup_v:
6422       case NEON::BI__builtin_neon_vld4_dup_v:
6423       case NEON::BI__builtin_neon_vld4q_dup_v:
6424         // Get the alignment for the argument in addition to the value;
6425         // we'll use it later.
6426         PtrOp1 = EmitPointerWithAlignment(E->getArg(1));
6427         Ops.push_back(PtrOp1.getPointer());
6428         continue;
6429       }
6430     }
6431 
6432     if ((ICEArguments & (1 << i)) == 0) {
6433       Ops.push_back(EmitScalarExpr(E->getArg(i)));
6434     } else {
6435       // If this is required to be a constant, constant fold it so that we know
6436       // that the generated intrinsic gets a ConstantInt.
6437       llvm::APSInt Result;
6438       bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
6439       assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
6440       Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
6441     }
6442   }
6443 
6444   switch (BuiltinID) {
6445   default: break;
6446 
6447   case NEON::BI__builtin_neon_vget_lane_i8:
6448   case NEON::BI__builtin_neon_vget_lane_i16:
6449   case NEON::BI__builtin_neon_vget_lane_i32:
6450   case NEON::BI__builtin_neon_vget_lane_i64:
6451   case NEON::BI__builtin_neon_vget_lane_f32:
6452   case NEON::BI__builtin_neon_vgetq_lane_i8:
6453   case NEON::BI__builtin_neon_vgetq_lane_i16:
6454   case NEON::BI__builtin_neon_vgetq_lane_i32:
6455   case NEON::BI__builtin_neon_vgetq_lane_i64:
6456   case NEON::BI__builtin_neon_vgetq_lane_f32:
6457     return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane");
6458 
6459   case NEON::BI__builtin_neon_vrndns_f32: {
6460     Value *Arg = EmitScalarExpr(E->getArg(0));
6461     llvm::Type *Tys[] = {Arg->getType()};
6462     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vrintn, Tys);
6463     return Builder.CreateCall(F, {Arg}, "vrndn"); }
6464 
6465   case NEON::BI__builtin_neon_vset_lane_i8:
6466   case NEON::BI__builtin_neon_vset_lane_i16:
6467   case NEON::BI__builtin_neon_vset_lane_i32:
6468   case NEON::BI__builtin_neon_vset_lane_i64:
6469   case NEON::BI__builtin_neon_vset_lane_f32:
6470   case NEON::BI__builtin_neon_vsetq_lane_i8:
6471   case NEON::BI__builtin_neon_vsetq_lane_i16:
6472   case NEON::BI__builtin_neon_vsetq_lane_i32:
6473   case NEON::BI__builtin_neon_vsetq_lane_i64:
6474   case NEON::BI__builtin_neon_vsetq_lane_f32:
6475     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
6476 
6477   case NEON::BI__builtin_neon_vsha1h_u32:
6478     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops,
6479                         "vsha1h");
6480   case NEON::BI__builtin_neon_vsha1cq_u32:
6481     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops,
6482                         "vsha1h");
6483   case NEON::BI__builtin_neon_vsha1pq_u32:
6484     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops,
6485                         "vsha1h");
6486   case NEON::BI__builtin_neon_vsha1mq_u32:
6487     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops,
6488                         "vsha1h");
6489 
6490   // The ARM _MoveToCoprocessor builtins put the input register value as
6491   // the first argument, but the LLVM intrinsic expects it as the third one.
6492   case ARM::BI_MoveToCoprocessor:
6493   case ARM::BI_MoveToCoprocessor2: {
6494     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ?
6495                                    Intrinsic::arm_mcr : Intrinsic::arm_mcr2);
6496     return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0],
6497                                   Ops[3], Ops[4], Ops[5]});
6498   }
6499   case ARM::BI_BitScanForward:
6500   case ARM::BI_BitScanForward64:
6501     return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E);
6502   case ARM::BI_BitScanReverse:
6503   case ARM::BI_BitScanReverse64:
6504     return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E);
6505 
6506   case ARM::BI_InterlockedAnd64:
6507     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E);
6508   case ARM::BI_InterlockedExchange64:
6509     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E);
6510   case ARM::BI_InterlockedExchangeAdd64:
6511     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E);
6512   case ARM::BI_InterlockedExchangeSub64:
6513     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E);
6514   case ARM::BI_InterlockedOr64:
6515     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E);
6516   case ARM::BI_InterlockedXor64:
6517     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E);
6518   case ARM::BI_InterlockedDecrement64:
6519     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E);
6520   case ARM::BI_InterlockedIncrement64:
6521     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E);
6522   case ARM::BI_InterlockedExchangeAdd8_acq:
6523   case ARM::BI_InterlockedExchangeAdd16_acq:
6524   case ARM::BI_InterlockedExchangeAdd_acq:
6525   case ARM::BI_InterlockedExchangeAdd64_acq:
6526     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_acq, E);
6527   case ARM::BI_InterlockedExchangeAdd8_rel:
6528   case ARM::BI_InterlockedExchangeAdd16_rel:
6529   case ARM::BI_InterlockedExchangeAdd_rel:
6530   case ARM::BI_InterlockedExchangeAdd64_rel:
6531     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_rel, E);
6532   case ARM::BI_InterlockedExchangeAdd8_nf:
6533   case ARM::BI_InterlockedExchangeAdd16_nf:
6534   case ARM::BI_InterlockedExchangeAdd_nf:
6535   case ARM::BI_InterlockedExchangeAdd64_nf:
6536     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_nf, E);
6537   case ARM::BI_InterlockedExchange8_acq:
6538   case ARM::BI_InterlockedExchange16_acq:
6539   case ARM::BI_InterlockedExchange_acq:
6540   case ARM::BI_InterlockedExchange64_acq:
6541     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_acq, E);
6542   case ARM::BI_InterlockedExchange8_rel:
6543   case ARM::BI_InterlockedExchange16_rel:
6544   case ARM::BI_InterlockedExchange_rel:
6545   case ARM::BI_InterlockedExchange64_rel:
6546     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_rel, E);
6547   case ARM::BI_InterlockedExchange8_nf:
6548   case ARM::BI_InterlockedExchange16_nf:
6549   case ARM::BI_InterlockedExchange_nf:
6550   case ARM::BI_InterlockedExchange64_nf:
6551     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_nf, E);
6552   case ARM::BI_InterlockedCompareExchange8_acq:
6553   case ARM::BI_InterlockedCompareExchange16_acq:
6554   case ARM::BI_InterlockedCompareExchange_acq:
6555   case ARM::BI_InterlockedCompareExchange64_acq:
6556     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_acq, E);
6557   case ARM::BI_InterlockedCompareExchange8_rel:
6558   case ARM::BI_InterlockedCompareExchange16_rel:
6559   case ARM::BI_InterlockedCompareExchange_rel:
6560   case ARM::BI_InterlockedCompareExchange64_rel:
6561     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_rel, E);
6562   case ARM::BI_InterlockedCompareExchange8_nf:
6563   case ARM::BI_InterlockedCompareExchange16_nf:
6564   case ARM::BI_InterlockedCompareExchange_nf:
6565   case ARM::BI_InterlockedCompareExchange64_nf:
6566     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_nf, E);
6567   case ARM::BI_InterlockedOr8_acq:
6568   case ARM::BI_InterlockedOr16_acq:
6569   case ARM::BI_InterlockedOr_acq:
6570   case ARM::BI_InterlockedOr64_acq:
6571     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_acq, E);
6572   case ARM::BI_InterlockedOr8_rel:
6573   case ARM::BI_InterlockedOr16_rel:
6574   case ARM::BI_InterlockedOr_rel:
6575   case ARM::BI_InterlockedOr64_rel:
6576     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_rel, E);
6577   case ARM::BI_InterlockedOr8_nf:
6578   case ARM::BI_InterlockedOr16_nf:
6579   case ARM::BI_InterlockedOr_nf:
6580   case ARM::BI_InterlockedOr64_nf:
6581     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_nf, E);
6582   case ARM::BI_InterlockedXor8_acq:
6583   case ARM::BI_InterlockedXor16_acq:
6584   case ARM::BI_InterlockedXor_acq:
6585   case ARM::BI_InterlockedXor64_acq:
6586     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_acq, E);
6587   case ARM::BI_InterlockedXor8_rel:
6588   case ARM::BI_InterlockedXor16_rel:
6589   case ARM::BI_InterlockedXor_rel:
6590   case ARM::BI_InterlockedXor64_rel:
6591     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_rel, E);
6592   case ARM::BI_InterlockedXor8_nf:
6593   case ARM::BI_InterlockedXor16_nf:
6594   case ARM::BI_InterlockedXor_nf:
6595   case ARM::BI_InterlockedXor64_nf:
6596     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_nf, E);
6597   case ARM::BI_InterlockedAnd8_acq:
6598   case ARM::BI_InterlockedAnd16_acq:
6599   case ARM::BI_InterlockedAnd_acq:
6600   case ARM::BI_InterlockedAnd64_acq:
6601     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_acq, E);
6602   case ARM::BI_InterlockedAnd8_rel:
6603   case ARM::BI_InterlockedAnd16_rel:
6604   case ARM::BI_InterlockedAnd_rel:
6605   case ARM::BI_InterlockedAnd64_rel:
6606     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_rel, E);
6607   case ARM::BI_InterlockedAnd8_nf:
6608   case ARM::BI_InterlockedAnd16_nf:
6609   case ARM::BI_InterlockedAnd_nf:
6610   case ARM::BI_InterlockedAnd64_nf:
6611     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_nf, E);
6612   case ARM::BI_InterlockedIncrement16_acq:
6613   case ARM::BI_InterlockedIncrement_acq:
6614   case ARM::BI_InterlockedIncrement64_acq:
6615     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_acq, E);
6616   case ARM::BI_InterlockedIncrement16_rel:
6617   case ARM::BI_InterlockedIncrement_rel:
6618   case ARM::BI_InterlockedIncrement64_rel:
6619     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_rel, E);
6620   case ARM::BI_InterlockedIncrement16_nf:
6621   case ARM::BI_InterlockedIncrement_nf:
6622   case ARM::BI_InterlockedIncrement64_nf:
6623     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_nf, E);
6624   case ARM::BI_InterlockedDecrement16_acq:
6625   case ARM::BI_InterlockedDecrement_acq:
6626   case ARM::BI_InterlockedDecrement64_acq:
6627     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_acq, E);
6628   case ARM::BI_InterlockedDecrement16_rel:
6629   case ARM::BI_InterlockedDecrement_rel:
6630   case ARM::BI_InterlockedDecrement64_rel:
6631     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_rel, E);
6632   case ARM::BI_InterlockedDecrement16_nf:
6633   case ARM::BI_InterlockedDecrement_nf:
6634   case ARM::BI_InterlockedDecrement64_nf:
6635     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_nf, E);
6636   }
6637 
6638   // Get the last argument, which specifies the vector type.
6639   assert(HasExtraArg);
6640   llvm::APSInt Result;
6641   const Expr *Arg = E->getArg(E->getNumArgs()-1);
6642   if (!Arg->isIntegerConstantExpr(Result, getContext()))
6643     return nullptr;
6644 
6645   if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
6646       BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
6647     // Determine the overloaded type of this builtin.
6648     llvm::Type *Ty;
6649     if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
6650       Ty = FloatTy;
6651     else
6652       Ty = DoubleTy;
6653 
6654     // Determine whether this is an unsigned conversion or not.
6655     bool usgn = Result.getZExtValue() == 1;
6656     unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
6657 
6658     // Call the appropriate intrinsic.
6659     Function *F = CGM.getIntrinsic(Int, Ty);
6660     return Builder.CreateCall(F, Ops, "vcvtr");
6661   }
6662 
6663   // Determine the type of this overloaded NEON intrinsic.
6664   NeonTypeFlags Type(Result.getZExtValue());
6665   bool usgn = Type.isUnsigned();
6666   bool rightShift = false;
6667 
6668   llvm::VectorType *VTy = GetNeonType(this, Type,
6669                                       getTarget().hasLegalHalfType());
6670   llvm::Type *Ty = VTy;
6671   if (!Ty)
6672     return nullptr;
6673 
6674   // Many NEON builtins have identical semantics and uses in ARM and
6675   // AArch64. Emit these in a single function.
6676   auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap);
6677   const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap(
6678       IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted);
6679   if (Builtin)
6680     return EmitCommonNeonBuiltinExpr(
6681         Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
6682         Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1, Arch);
6683 
6684   unsigned Int;
6685   switch (BuiltinID) {
6686   default: return nullptr;
6687   case NEON::BI__builtin_neon_vld1q_lane_v:
6688     // Handle 64-bit integer elements as a special case.  Use shuffles of
6689     // one-element vectors to avoid poor code for i64 in the backend.
6690     if (VTy->getElementType()->isIntegerTy(64)) {
6691       // Extract the other lane.
6692       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6693       uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
6694       Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
6695       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
6696       // Load the value as a one-element vector.
6697       Ty = llvm::VectorType::get(VTy->getElementType(), 1);
6698       llvm::Type *Tys[] = {Ty, Int8PtrTy};
6699       Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys);
6700       Value *Align = getAlignmentValue32(PtrOp0);
6701       Value *Ld = Builder.CreateCall(F, {Ops[0], Align});
6702       // Combine them.
6703       uint32_t Indices[] = {1 - Lane, Lane};
6704       SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices);
6705       return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane");
6706     }
6707     LLVM_FALLTHROUGH;
6708   case NEON::BI__builtin_neon_vld1_lane_v: {
6709     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6710     PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType());
6711     Value *Ld = Builder.CreateLoad(PtrOp0);
6712     return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
6713   }
6714   case NEON::BI__builtin_neon_vqrshrn_n_v:
6715     Int =
6716       usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
6717     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
6718                         1, true);
6719   case NEON::BI__builtin_neon_vqrshrun_n_v:
6720     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
6721                         Ops, "vqrshrun_n", 1, true);
6722   case NEON::BI__builtin_neon_vqshrn_n_v:
6723     Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
6724     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
6725                         1, true);
6726   case NEON::BI__builtin_neon_vqshrun_n_v:
6727     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
6728                         Ops, "vqshrun_n", 1, true);
6729   case NEON::BI__builtin_neon_vrecpe_v:
6730   case NEON::BI__builtin_neon_vrecpeq_v:
6731     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
6732                         Ops, "vrecpe");
6733   case NEON::BI__builtin_neon_vrshrn_n_v:
6734     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
6735                         Ops, "vrshrn_n", 1, true);
6736   case NEON::BI__builtin_neon_vrsra_n_v:
6737   case NEON::BI__builtin_neon_vrsraq_n_v:
6738     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6739     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6740     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
6741     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
6742     Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]});
6743     return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
6744   case NEON::BI__builtin_neon_vsri_n_v:
6745   case NEON::BI__builtin_neon_vsriq_n_v:
6746     rightShift = true;
6747     LLVM_FALLTHROUGH;
6748   case NEON::BI__builtin_neon_vsli_n_v:
6749   case NEON::BI__builtin_neon_vsliq_n_v:
6750     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
6751     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
6752                         Ops, "vsli_n");
6753   case NEON::BI__builtin_neon_vsra_n_v:
6754   case NEON::BI__builtin_neon_vsraq_n_v:
6755     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6756     Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
6757     return Builder.CreateAdd(Ops[0], Ops[1]);
6758   case NEON::BI__builtin_neon_vst1q_lane_v:
6759     // Handle 64-bit integer elements as a special case.  Use a shuffle to get
6760     // a one-element vector and avoid poor code for i64 in the backend.
6761     if (VTy->getElementType()->isIntegerTy(64)) {
6762       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6763       Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
6764       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
6765       Ops[2] = getAlignmentValue32(PtrOp0);
6766       llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()};
6767       return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
6768                                                  Tys), Ops);
6769     }
6770     LLVM_FALLTHROUGH;
6771   case NEON::BI__builtin_neon_vst1_lane_v: {
6772     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6773     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
6774     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
6775     auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty));
6776     return St;
6777   }
6778   case NEON::BI__builtin_neon_vtbl1_v:
6779     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
6780                         Ops, "vtbl1");
6781   case NEON::BI__builtin_neon_vtbl2_v:
6782     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
6783                         Ops, "vtbl2");
6784   case NEON::BI__builtin_neon_vtbl3_v:
6785     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
6786                         Ops, "vtbl3");
6787   case NEON::BI__builtin_neon_vtbl4_v:
6788     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
6789                         Ops, "vtbl4");
6790   case NEON::BI__builtin_neon_vtbx1_v:
6791     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
6792                         Ops, "vtbx1");
6793   case NEON::BI__builtin_neon_vtbx2_v:
6794     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
6795                         Ops, "vtbx2");
6796   case NEON::BI__builtin_neon_vtbx3_v:
6797     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
6798                         Ops, "vtbx3");
6799   case NEON::BI__builtin_neon_vtbx4_v:
6800     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
6801                         Ops, "vtbx4");
6802   }
6803 }
6804 
6805 static llvm::Value *SignOrZeroExtend(CGBuilderTy &Builder, llvm::Value *V,
6806                                      llvm::Type *T, bool Unsigned) {
6807   // Helper function called by Tablegen-constructed ARM MVE builtin codegen,
6808   // which finds it convenient to specify signed/unsigned as a boolean flag.
6809   return Unsigned ? Builder.CreateZExt(V, T) : Builder.CreateSExt(V, T);
6810 }
6811 
6812 static llvm::Value *ARMMVEVectorSplat(CGBuilderTy &Builder, llvm::Value *V) {
6813   // MVE-specific helper function for a vector splat, which infers the element
6814   // count of the output vector by knowing that MVE vectors are all 128 bits
6815   // wide.
6816   unsigned Elements = 128 / V->getType()->getPrimitiveSizeInBits();
6817   return Builder.CreateVectorSplat(Elements, V);
6818 }
6819 
6820 Value *CodeGenFunction::EmitARMMVEBuiltinExpr(unsigned BuiltinID,
6821                                               const CallExpr *E,
6822                                               ReturnValueSlot ReturnValue,
6823                                               llvm::Triple::ArchType Arch) {
6824   enum class CustomCodeGen { VLD24, VST24 } CustomCodeGenType;
6825   Intrinsic::ID IRIntr;
6826   unsigned NumVectors;
6827 
6828   // Code autogenerated by Tablegen will handle all the simple builtins.
6829   switch (BuiltinID) {
6830     #include "clang/Basic/arm_mve_builtin_cg.inc"
6831 
6832     // If we didn't match an MVE builtin id at all, go back to the
6833     // main EmitARMBuiltinExpr.
6834   default:
6835     return nullptr;
6836   }
6837 
6838   // Anything that breaks from that switch is an MVE builtin that
6839   // needs handwritten code to generate.
6840 
6841   switch (CustomCodeGenType) {
6842 
6843   case CustomCodeGen::VLD24: {
6844     llvm::SmallVector<Value *, 4> Ops;
6845     llvm::SmallVector<llvm::Type *, 4> Tys;
6846 
6847     auto MvecCType = E->getType();
6848     auto MvecLType = ConvertType(MvecCType);
6849     assert(MvecLType->isStructTy() &&
6850            "Return type for vld[24]q should be a struct");
6851     assert(MvecLType->getStructNumElements() == 1 &&
6852            "Return-type struct for vld[24]q should have one element");
6853     auto MvecLTypeInner = MvecLType->getStructElementType(0);
6854     assert(MvecLTypeInner->isArrayTy() &&
6855            "Return-type struct for vld[24]q should contain an array");
6856     assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&
6857            "Array member of return-type struct vld[24]q has wrong length");
6858     auto VecLType = MvecLTypeInner->getArrayElementType();
6859 
6860     Tys.push_back(VecLType);
6861 
6862     auto Addr = E->getArg(0);
6863     Ops.push_back(EmitScalarExpr(Addr));
6864     Tys.push_back(ConvertType(Addr->getType()));
6865 
6866     Function *F = CGM.getIntrinsic(IRIntr, makeArrayRef(Tys));
6867     Value *LoadResult = Builder.CreateCall(F, Ops);
6868     Value *MvecOut = UndefValue::get(MvecLType);
6869     for (unsigned i = 0; i < NumVectors; ++i) {
6870       Value *Vec = Builder.CreateExtractValue(LoadResult, i);
6871       MvecOut = Builder.CreateInsertValue(MvecOut, Vec, {0, i});
6872     }
6873 
6874     if (ReturnValue.isNull())
6875       return MvecOut;
6876     else
6877       return Builder.CreateStore(MvecOut, ReturnValue.getValue());
6878   }
6879 
6880   case CustomCodeGen::VST24: {
6881     llvm::SmallVector<Value *, 4> Ops;
6882     llvm::SmallVector<llvm::Type *, 4> Tys;
6883 
6884     auto Addr = E->getArg(0);
6885     Ops.push_back(EmitScalarExpr(Addr));
6886     Tys.push_back(ConvertType(Addr->getType()));
6887 
6888     auto MvecCType = E->getArg(1)->getType();
6889     auto MvecLType = ConvertType(MvecCType);
6890     assert(MvecLType->isStructTy() && "Data type for vst2q should be a struct");
6891     assert(MvecLType->getStructNumElements() == 1 &&
6892            "Data-type struct for vst2q should have one element");
6893     auto MvecLTypeInner = MvecLType->getStructElementType(0);
6894     assert(MvecLTypeInner->isArrayTy() &&
6895            "Data-type struct for vst2q should contain an array");
6896     assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&
6897            "Array member of return-type struct vld[24]q has wrong length");
6898     auto VecLType = MvecLTypeInner->getArrayElementType();
6899 
6900     Tys.push_back(VecLType);
6901 
6902     AggValueSlot MvecSlot = CreateAggTemp(MvecCType);
6903     EmitAggExpr(E->getArg(1), MvecSlot);
6904     auto Mvec = Builder.CreateLoad(MvecSlot.getAddress());
6905     for (unsigned i = 0; i < NumVectors; i++)
6906       Ops.push_back(Builder.CreateExtractValue(Mvec, {0, i}));
6907 
6908     Function *F = CGM.getIntrinsic(IRIntr, makeArrayRef(Tys));
6909     Value *ToReturn = nullptr;
6910     for (unsigned i = 0; i < NumVectors; i++) {
6911       Ops.push_back(llvm::ConstantInt::get(Int32Ty, i));
6912       ToReturn = Builder.CreateCall(F, Ops);
6913       Ops.pop_back();
6914     }
6915     return ToReturn;
6916   }
6917   }
6918   llvm_unreachable("unknown custom codegen type.");
6919 }
6920 
6921 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID,
6922                                       const CallExpr *E,
6923                                       SmallVectorImpl<Value *> &Ops,
6924                                       llvm::Triple::ArchType Arch) {
6925   unsigned int Int = 0;
6926   const char *s = nullptr;
6927 
6928   switch (BuiltinID) {
6929   default:
6930     return nullptr;
6931   case NEON::BI__builtin_neon_vtbl1_v:
6932   case NEON::BI__builtin_neon_vqtbl1_v:
6933   case NEON::BI__builtin_neon_vqtbl1q_v:
6934   case NEON::BI__builtin_neon_vtbl2_v:
6935   case NEON::BI__builtin_neon_vqtbl2_v:
6936   case NEON::BI__builtin_neon_vqtbl2q_v:
6937   case NEON::BI__builtin_neon_vtbl3_v:
6938   case NEON::BI__builtin_neon_vqtbl3_v:
6939   case NEON::BI__builtin_neon_vqtbl3q_v:
6940   case NEON::BI__builtin_neon_vtbl4_v:
6941   case NEON::BI__builtin_neon_vqtbl4_v:
6942   case NEON::BI__builtin_neon_vqtbl4q_v:
6943     break;
6944   case NEON::BI__builtin_neon_vtbx1_v:
6945   case NEON::BI__builtin_neon_vqtbx1_v:
6946   case NEON::BI__builtin_neon_vqtbx1q_v:
6947   case NEON::BI__builtin_neon_vtbx2_v:
6948   case NEON::BI__builtin_neon_vqtbx2_v:
6949   case NEON::BI__builtin_neon_vqtbx2q_v:
6950   case NEON::BI__builtin_neon_vtbx3_v:
6951   case NEON::BI__builtin_neon_vqtbx3_v:
6952   case NEON::BI__builtin_neon_vqtbx3q_v:
6953   case NEON::BI__builtin_neon_vtbx4_v:
6954   case NEON::BI__builtin_neon_vqtbx4_v:
6955   case NEON::BI__builtin_neon_vqtbx4q_v:
6956     break;
6957   }
6958 
6959   assert(E->getNumArgs() >= 3);
6960 
6961   // Get the last argument, which specifies the vector type.
6962   llvm::APSInt Result;
6963   const Expr *Arg = E->getArg(E->getNumArgs() - 1);
6964   if (!Arg->isIntegerConstantExpr(Result, CGF.getContext()))
6965     return nullptr;
6966 
6967   // Determine the type of this overloaded NEON intrinsic.
6968   NeonTypeFlags Type(Result.getZExtValue());
6969   llvm::VectorType *Ty = GetNeonType(&CGF, Type);
6970   if (!Ty)
6971     return nullptr;
6972 
6973   CodeGen::CGBuilderTy &Builder = CGF.Builder;
6974 
6975   // AArch64 scalar builtins are not overloaded, they do not have an extra
6976   // argument that specifies the vector type, need to handle each case.
6977   switch (BuiltinID) {
6978   case NEON::BI__builtin_neon_vtbl1_v: {
6979     return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr,
6980                               Ops[1], Ty, Intrinsic::aarch64_neon_tbl1,
6981                               "vtbl1");
6982   }
6983   case NEON::BI__builtin_neon_vtbl2_v: {
6984     return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr,
6985                               Ops[2], Ty, Intrinsic::aarch64_neon_tbl1,
6986                               "vtbl1");
6987   }
6988   case NEON::BI__builtin_neon_vtbl3_v: {
6989     return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr,
6990                               Ops[3], Ty, Intrinsic::aarch64_neon_tbl2,
6991                               "vtbl2");
6992   }
6993   case NEON::BI__builtin_neon_vtbl4_v: {
6994     return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr,
6995                               Ops[4], Ty, Intrinsic::aarch64_neon_tbl2,
6996                               "vtbl2");
6997   }
6998   case NEON::BI__builtin_neon_vtbx1_v: {
6999     Value *TblRes =
7000         packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2],
7001                            Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1");
7002 
7003     llvm::Constant *EightV = ConstantInt::get(Ty, 8);
7004     Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV);
7005     CmpRes = Builder.CreateSExt(CmpRes, Ty);
7006 
7007     Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
7008     Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
7009     return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
7010   }
7011   case NEON::BI__builtin_neon_vtbx2_v: {
7012     return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0],
7013                               Ops[3], Ty, Intrinsic::aarch64_neon_tbx1,
7014                               "vtbx1");
7015   }
7016   case NEON::BI__builtin_neon_vtbx3_v: {
7017     Value *TblRes =
7018         packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4],
7019                            Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2");
7020 
7021     llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24);
7022     Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4],
7023                                            TwentyFourV);
7024     CmpRes = Builder.CreateSExt(CmpRes, Ty);
7025 
7026     Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
7027     Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
7028     return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
7029   }
7030   case NEON::BI__builtin_neon_vtbx4_v: {
7031     return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0],
7032                               Ops[5], Ty, Intrinsic::aarch64_neon_tbx2,
7033                               "vtbx2");
7034   }
7035   case NEON::BI__builtin_neon_vqtbl1_v:
7036   case NEON::BI__builtin_neon_vqtbl1q_v:
7037     Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break;
7038   case NEON::BI__builtin_neon_vqtbl2_v:
7039   case NEON::BI__builtin_neon_vqtbl2q_v: {
7040     Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break;
7041   case NEON::BI__builtin_neon_vqtbl3_v:
7042   case NEON::BI__builtin_neon_vqtbl3q_v:
7043     Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break;
7044   case NEON::BI__builtin_neon_vqtbl4_v:
7045   case NEON::BI__builtin_neon_vqtbl4q_v:
7046     Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break;
7047   case NEON::BI__builtin_neon_vqtbx1_v:
7048   case NEON::BI__builtin_neon_vqtbx1q_v:
7049     Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break;
7050   case NEON::BI__builtin_neon_vqtbx2_v:
7051   case NEON::BI__builtin_neon_vqtbx2q_v:
7052     Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break;
7053   case NEON::BI__builtin_neon_vqtbx3_v:
7054   case NEON::BI__builtin_neon_vqtbx3q_v:
7055     Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break;
7056   case NEON::BI__builtin_neon_vqtbx4_v:
7057   case NEON::BI__builtin_neon_vqtbx4q_v:
7058     Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break;
7059   }
7060   }
7061 
7062   if (!Int)
7063     return nullptr;
7064 
7065   Function *F = CGF.CGM.getIntrinsic(Int, Ty);
7066   return CGF.EmitNeonCall(F, Ops, s);
7067 }
7068 
7069 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) {
7070   llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
7071   Op = Builder.CreateBitCast(Op, Int16Ty);
7072   Value *V = UndefValue::get(VTy);
7073   llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
7074   Op = Builder.CreateInsertElement(V, Op, CI);
7075   return Op;
7076 }
7077 
7078 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
7079                                                const CallExpr *E,
7080                                                llvm::Triple::ArchType Arch) {
7081   unsigned HintID = static_cast<unsigned>(-1);
7082   switch (BuiltinID) {
7083   default: break;
7084   case AArch64::BI__builtin_arm_nop:
7085     HintID = 0;
7086     break;
7087   case AArch64::BI__builtin_arm_yield:
7088   case AArch64::BI__yield:
7089     HintID = 1;
7090     break;
7091   case AArch64::BI__builtin_arm_wfe:
7092   case AArch64::BI__wfe:
7093     HintID = 2;
7094     break;
7095   case AArch64::BI__builtin_arm_wfi:
7096   case AArch64::BI__wfi:
7097     HintID = 3;
7098     break;
7099   case AArch64::BI__builtin_arm_sev:
7100   case AArch64::BI__sev:
7101     HintID = 4;
7102     break;
7103   case AArch64::BI__builtin_arm_sevl:
7104   case AArch64::BI__sevl:
7105     HintID = 5;
7106     break;
7107   }
7108 
7109   if (HintID != static_cast<unsigned>(-1)) {
7110     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint);
7111     return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID));
7112   }
7113 
7114   if (BuiltinID == AArch64::BI__builtin_arm_prefetch) {
7115     Value *Address         = EmitScalarExpr(E->getArg(0));
7116     Value *RW              = EmitScalarExpr(E->getArg(1));
7117     Value *CacheLevel      = EmitScalarExpr(E->getArg(2));
7118     Value *RetentionPolicy = EmitScalarExpr(E->getArg(3));
7119     Value *IsData          = EmitScalarExpr(E->getArg(4));
7120 
7121     Value *Locality = nullptr;
7122     if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) {
7123       // Temporal fetch, needs to convert cache level to locality.
7124       Locality = llvm::ConstantInt::get(Int32Ty,
7125         -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3);
7126     } else {
7127       // Streaming fetch.
7128       Locality = llvm::ConstantInt::get(Int32Ty, 0);
7129     }
7130 
7131     // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify
7132     // PLDL3STRM or PLDL2STRM.
7133     Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
7134     return Builder.CreateCall(F, {Address, RW, Locality, IsData});
7135   }
7136 
7137   if (BuiltinID == AArch64::BI__builtin_arm_rbit) {
7138     assert((getContext().getTypeSize(E->getType()) == 32) &&
7139            "rbit of unusual size!");
7140     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7141     return Builder.CreateCall(
7142         CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
7143   }
7144   if (BuiltinID == AArch64::BI__builtin_arm_rbit64) {
7145     assert((getContext().getTypeSize(E->getType()) == 64) &&
7146            "rbit of unusual size!");
7147     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7148     return Builder.CreateCall(
7149         CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
7150   }
7151 
7152   if (BuiltinID == AArch64::BI__builtin_arm_cls) {
7153     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7154     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls), Arg,
7155                               "cls");
7156   }
7157   if (BuiltinID == AArch64::BI__builtin_arm_cls64) {
7158     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7159     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls64), Arg,
7160                               "cls");
7161   }
7162 
7163   if (BuiltinID == AArch64::BI__builtin_arm_jcvt) {
7164     assert((getContext().getTypeSize(E->getType()) == 32) &&
7165            "__jcvt of unusual size!");
7166     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7167     return Builder.CreateCall(
7168         CGM.getIntrinsic(Intrinsic::aarch64_fjcvtzs), Arg);
7169   }
7170 
7171   if (BuiltinID == AArch64::BI__clear_cache) {
7172     assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
7173     const FunctionDecl *FD = E->getDirectCallee();
7174     Value *Ops[2];
7175     for (unsigned i = 0; i < 2; i++)
7176       Ops[i] = EmitScalarExpr(E->getArg(i));
7177     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
7178     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
7179     StringRef Name = FD->getName();
7180     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
7181   }
7182 
7183   if ((BuiltinID == AArch64::BI__builtin_arm_ldrex ||
7184       BuiltinID == AArch64::BI__builtin_arm_ldaex) &&
7185       getContext().getTypeSize(E->getType()) == 128) {
7186     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
7187                                        ? Intrinsic::aarch64_ldaxp
7188                                        : Intrinsic::aarch64_ldxp);
7189 
7190     Value *LdPtr = EmitScalarExpr(E->getArg(0));
7191     Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
7192                                     "ldxp");
7193 
7194     Value *Val0 = Builder.CreateExtractValue(Val, 1);
7195     Value *Val1 = Builder.CreateExtractValue(Val, 0);
7196     llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
7197     Val0 = Builder.CreateZExt(Val0, Int128Ty);
7198     Val1 = Builder.CreateZExt(Val1, Int128Ty);
7199 
7200     Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64);
7201     Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
7202     Val = Builder.CreateOr(Val, Val1);
7203     return Builder.CreateBitCast(Val, ConvertType(E->getType()));
7204   } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex ||
7205              BuiltinID == AArch64::BI__builtin_arm_ldaex) {
7206     Value *LoadAddr = EmitScalarExpr(E->getArg(0));
7207 
7208     QualType Ty = E->getType();
7209     llvm::Type *RealResTy = ConvertType(Ty);
7210     llvm::Type *PtrTy = llvm::IntegerType::get(
7211         getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo();
7212     LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy);
7213 
7214     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
7215                                        ? Intrinsic::aarch64_ldaxr
7216                                        : Intrinsic::aarch64_ldxr,
7217                                    PtrTy);
7218     Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr");
7219 
7220     if (RealResTy->isPointerTy())
7221       return Builder.CreateIntToPtr(Val, RealResTy);
7222 
7223     llvm::Type *IntResTy = llvm::IntegerType::get(
7224         getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy));
7225     Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
7226     return Builder.CreateBitCast(Val, RealResTy);
7227   }
7228 
7229   if ((BuiltinID == AArch64::BI__builtin_arm_strex ||
7230        BuiltinID == AArch64::BI__builtin_arm_stlex) &&
7231       getContext().getTypeSize(E->getArg(0)->getType()) == 128) {
7232     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
7233                                        ? Intrinsic::aarch64_stlxp
7234                                        : Intrinsic::aarch64_stxp);
7235     llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty);
7236 
7237     Address Tmp = CreateMemTemp(E->getArg(0)->getType());
7238     EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true);
7239 
7240     Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy));
7241     llvm::Value *Val = Builder.CreateLoad(Tmp);
7242 
7243     Value *Arg0 = Builder.CreateExtractValue(Val, 0);
7244     Value *Arg1 = Builder.CreateExtractValue(Val, 1);
7245     Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)),
7246                                          Int8PtrTy);
7247     return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp");
7248   }
7249 
7250   if (BuiltinID == AArch64::BI__builtin_arm_strex ||
7251       BuiltinID == AArch64::BI__builtin_arm_stlex) {
7252     Value *StoreVal = EmitScalarExpr(E->getArg(0));
7253     Value *StoreAddr = EmitScalarExpr(E->getArg(1));
7254 
7255     QualType Ty = E->getArg(0)->getType();
7256     llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
7257                                                  getContext().getTypeSize(Ty));
7258     StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
7259 
7260     if (StoreVal->getType()->isPointerTy())
7261       StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty);
7262     else {
7263       llvm::Type *IntTy = llvm::IntegerType::get(
7264           getLLVMContext(),
7265           CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType()));
7266       StoreVal = Builder.CreateBitCast(StoreVal, IntTy);
7267       StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty);
7268     }
7269 
7270     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
7271                                        ? Intrinsic::aarch64_stlxr
7272                                        : Intrinsic::aarch64_stxr,
7273                                    StoreAddr->getType());
7274     return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr");
7275   }
7276 
7277   if (BuiltinID == AArch64::BI__getReg) {
7278     Expr::EvalResult Result;
7279     if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
7280       llvm_unreachable("Sema will ensure that the parameter is constant");
7281 
7282     llvm::APSInt Value = Result.Val.getInt();
7283     LLVMContext &Context = CGM.getLLVMContext();
7284     std::string Reg = Value == 31 ? "sp" : "x" + Value.toString(10);
7285 
7286     llvm::Metadata *Ops[] = {llvm::MDString::get(Context, Reg)};
7287     llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
7288     llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
7289 
7290     llvm::Function *F =
7291         CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty});
7292     return Builder.CreateCall(F, Metadata);
7293   }
7294 
7295   if (BuiltinID == AArch64::BI__builtin_arm_clrex) {
7296     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex);
7297     return Builder.CreateCall(F);
7298   }
7299 
7300   if (BuiltinID == AArch64::BI_ReadWriteBarrier)
7301     return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
7302                                llvm::SyncScope::SingleThread);
7303 
7304   // CRC32
7305   Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
7306   switch (BuiltinID) {
7307   case AArch64::BI__builtin_arm_crc32b:
7308     CRCIntrinsicID = Intrinsic::aarch64_crc32b; break;
7309   case AArch64::BI__builtin_arm_crc32cb:
7310     CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break;
7311   case AArch64::BI__builtin_arm_crc32h:
7312     CRCIntrinsicID = Intrinsic::aarch64_crc32h; break;
7313   case AArch64::BI__builtin_arm_crc32ch:
7314     CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break;
7315   case AArch64::BI__builtin_arm_crc32w:
7316     CRCIntrinsicID = Intrinsic::aarch64_crc32w; break;
7317   case AArch64::BI__builtin_arm_crc32cw:
7318     CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break;
7319   case AArch64::BI__builtin_arm_crc32d:
7320     CRCIntrinsicID = Intrinsic::aarch64_crc32x; break;
7321   case AArch64::BI__builtin_arm_crc32cd:
7322     CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break;
7323   }
7324 
7325   if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
7326     Value *Arg0 = EmitScalarExpr(E->getArg(0));
7327     Value *Arg1 = EmitScalarExpr(E->getArg(1));
7328     Function *F = CGM.getIntrinsic(CRCIntrinsicID);
7329 
7330     llvm::Type *DataTy = F->getFunctionType()->getParamType(1);
7331     Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy);
7332 
7333     return Builder.CreateCall(F, {Arg0, Arg1});
7334   }
7335 
7336   // Memory Tagging Extensions (MTE) Intrinsics
7337   Intrinsic::ID MTEIntrinsicID = Intrinsic::not_intrinsic;
7338   switch (BuiltinID) {
7339   case AArch64::BI__builtin_arm_irg:
7340     MTEIntrinsicID = Intrinsic::aarch64_irg; break;
7341   case  AArch64::BI__builtin_arm_addg:
7342     MTEIntrinsicID = Intrinsic::aarch64_addg; break;
7343   case  AArch64::BI__builtin_arm_gmi:
7344     MTEIntrinsicID = Intrinsic::aarch64_gmi; break;
7345   case  AArch64::BI__builtin_arm_ldg:
7346     MTEIntrinsicID = Intrinsic::aarch64_ldg; break;
7347   case AArch64::BI__builtin_arm_stg:
7348     MTEIntrinsicID = Intrinsic::aarch64_stg; break;
7349   case AArch64::BI__builtin_arm_subp:
7350     MTEIntrinsicID = Intrinsic::aarch64_subp; break;
7351   }
7352 
7353   if (MTEIntrinsicID != Intrinsic::not_intrinsic) {
7354     llvm::Type *T = ConvertType(E->getType());
7355 
7356     if (MTEIntrinsicID == Intrinsic::aarch64_irg) {
7357       Value *Pointer = EmitScalarExpr(E->getArg(0));
7358       Value *Mask = EmitScalarExpr(E->getArg(1));
7359 
7360       Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
7361       Mask = Builder.CreateZExt(Mask, Int64Ty);
7362       Value *RV = Builder.CreateCall(
7363                        CGM.getIntrinsic(MTEIntrinsicID), {Pointer, Mask});
7364        return Builder.CreatePointerCast(RV, T);
7365     }
7366     if (MTEIntrinsicID == Intrinsic::aarch64_addg) {
7367       Value *Pointer = EmitScalarExpr(E->getArg(0));
7368       Value *TagOffset = EmitScalarExpr(E->getArg(1));
7369 
7370       Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
7371       TagOffset = Builder.CreateZExt(TagOffset, Int64Ty);
7372       Value *RV = Builder.CreateCall(
7373                        CGM.getIntrinsic(MTEIntrinsicID), {Pointer, TagOffset});
7374       return Builder.CreatePointerCast(RV, T);
7375     }
7376     if (MTEIntrinsicID == Intrinsic::aarch64_gmi) {
7377       Value *Pointer = EmitScalarExpr(E->getArg(0));
7378       Value *ExcludedMask = EmitScalarExpr(E->getArg(1));
7379 
7380       ExcludedMask = Builder.CreateZExt(ExcludedMask, Int64Ty);
7381       Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
7382       return Builder.CreateCall(
7383                        CGM.getIntrinsic(MTEIntrinsicID), {Pointer, ExcludedMask});
7384     }
7385     // Although it is possible to supply a different return
7386     // address (first arg) to this intrinsic, for now we set
7387     // return address same as input address.
7388     if (MTEIntrinsicID == Intrinsic::aarch64_ldg) {
7389       Value *TagAddress = EmitScalarExpr(E->getArg(0));
7390       TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy);
7391       Value *RV = Builder.CreateCall(
7392                     CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress});
7393       return Builder.CreatePointerCast(RV, T);
7394     }
7395     // Although it is possible to supply a different tag (to set)
7396     // to this intrinsic (as first arg), for now we supply
7397     // the tag that is in input address arg (common use case).
7398     if (MTEIntrinsicID == Intrinsic::aarch64_stg) {
7399         Value *TagAddress = EmitScalarExpr(E->getArg(0));
7400         TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy);
7401         return Builder.CreateCall(
7402                  CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress});
7403     }
7404     if (MTEIntrinsicID == Intrinsic::aarch64_subp) {
7405       Value *PointerA = EmitScalarExpr(E->getArg(0));
7406       Value *PointerB = EmitScalarExpr(E->getArg(1));
7407       PointerA = Builder.CreatePointerCast(PointerA, Int8PtrTy);
7408       PointerB = Builder.CreatePointerCast(PointerB, Int8PtrTy);
7409       return Builder.CreateCall(
7410                        CGM.getIntrinsic(MTEIntrinsicID), {PointerA, PointerB});
7411     }
7412   }
7413 
7414   if (BuiltinID == AArch64::BI__builtin_arm_rsr ||
7415       BuiltinID == AArch64::BI__builtin_arm_rsr64 ||
7416       BuiltinID == AArch64::BI__builtin_arm_rsrp ||
7417       BuiltinID == AArch64::BI__builtin_arm_wsr ||
7418       BuiltinID == AArch64::BI__builtin_arm_wsr64 ||
7419       BuiltinID == AArch64::BI__builtin_arm_wsrp) {
7420 
7421     bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr ||
7422                   BuiltinID == AArch64::BI__builtin_arm_rsr64 ||
7423                   BuiltinID == AArch64::BI__builtin_arm_rsrp;
7424 
7425     bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp ||
7426                             BuiltinID == AArch64::BI__builtin_arm_wsrp;
7427 
7428     bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr &&
7429                    BuiltinID != AArch64::BI__builtin_arm_wsr;
7430 
7431     llvm::Type *ValueType;
7432     llvm::Type *RegisterType = Int64Ty;
7433     if (IsPointerBuiltin) {
7434       ValueType = VoidPtrTy;
7435     } else if (Is64Bit) {
7436       ValueType = Int64Ty;
7437     } else {
7438       ValueType = Int32Ty;
7439     }
7440 
7441     return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead);
7442   }
7443 
7444   if (BuiltinID == AArch64::BI_ReadStatusReg ||
7445       BuiltinID == AArch64::BI_WriteStatusReg) {
7446     LLVMContext &Context = CGM.getLLVMContext();
7447 
7448     unsigned SysReg =
7449       E->getArg(0)->EvaluateKnownConstInt(getContext()).getZExtValue();
7450 
7451     std::string SysRegStr;
7452     llvm::raw_string_ostream(SysRegStr) <<
7453                        ((1 << 1) | ((SysReg >> 14) & 1))  << ":" <<
7454                        ((SysReg >> 11) & 7)               << ":" <<
7455                        ((SysReg >> 7)  & 15)              << ":" <<
7456                        ((SysReg >> 3)  & 15)              << ":" <<
7457                        ( SysReg        & 7);
7458 
7459     llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysRegStr) };
7460     llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
7461     llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
7462 
7463     llvm::Type *RegisterType = Int64Ty;
7464     llvm::Type *Types[] = { RegisterType };
7465 
7466     if (BuiltinID == AArch64::BI_ReadStatusReg) {
7467       llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types);
7468 
7469       return Builder.CreateCall(F, Metadata);
7470     }
7471 
7472     llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types);
7473     llvm::Value *ArgValue = EmitScalarExpr(E->getArg(1));
7474 
7475     return Builder.CreateCall(F, { Metadata, ArgValue });
7476   }
7477 
7478   if (BuiltinID == AArch64::BI_AddressOfReturnAddress) {
7479     llvm::Function *F =
7480         CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy);
7481     return Builder.CreateCall(F);
7482   }
7483 
7484   if (BuiltinID == AArch64::BI__builtin_sponentry) {
7485     llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry, AllocaInt8PtrTy);
7486     return Builder.CreateCall(F);
7487   }
7488 
7489   // Find out if any arguments are required to be integer constant
7490   // expressions.
7491   unsigned ICEArguments = 0;
7492   ASTContext::GetBuiltinTypeError Error;
7493   getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
7494   assert(Error == ASTContext::GE_None && "Should not codegen an error");
7495 
7496   llvm::SmallVector<Value*, 4> Ops;
7497   for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
7498     if ((ICEArguments & (1 << i)) == 0) {
7499       Ops.push_back(EmitScalarExpr(E->getArg(i)));
7500     } else {
7501       // If this is required to be a constant, constant fold it so that we know
7502       // that the generated intrinsic gets a ConstantInt.
7503       llvm::APSInt Result;
7504       bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
7505       assert(IsConst && "Constant arg isn't actually constant?");
7506       (void)IsConst;
7507       Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
7508     }
7509   }
7510 
7511   auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap);
7512   const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap(
7513       SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted);
7514 
7515   if (Builtin) {
7516     Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1)));
7517     Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E);
7518     assert(Result && "SISD intrinsic should have been handled");
7519     return Result;
7520   }
7521 
7522   llvm::APSInt Result;
7523   const Expr *Arg = E->getArg(E->getNumArgs()-1);
7524   NeonTypeFlags Type(0);
7525   if (Arg->isIntegerConstantExpr(Result, getContext()))
7526     // Determine the type of this overloaded NEON intrinsic.
7527     Type = NeonTypeFlags(Result.getZExtValue());
7528 
7529   bool usgn = Type.isUnsigned();
7530   bool quad = Type.isQuad();
7531 
7532   // Handle non-overloaded intrinsics first.
7533   switch (BuiltinID) {
7534   default: break;
7535   case NEON::BI__builtin_neon_vabsh_f16:
7536     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7537     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, HalfTy), Ops, "vabs");
7538   case NEON::BI__builtin_neon_vldrq_p128: {
7539     llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128);
7540     llvm::Type *Int128PTy = llvm::PointerType::get(Int128Ty, 0);
7541     Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy);
7542     return Builder.CreateAlignedLoad(Int128Ty, Ptr,
7543                                      CharUnits::fromQuantity(16));
7544   }
7545   case NEON::BI__builtin_neon_vstrq_p128: {
7546     llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
7547     Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy);
7548     return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr);
7549   }
7550   case NEON::BI__builtin_neon_vcvts_u32_f32:
7551   case NEON::BI__builtin_neon_vcvtd_u64_f64:
7552     usgn = true;
7553     LLVM_FALLTHROUGH;
7554   case NEON::BI__builtin_neon_vcvts_s32_f32:
7555   case NEON::BI__builtin_neon_vcvtd_s64_f64: {
7556     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7557     bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
7558     llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
7559     llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
7560     Ops[0] = Builder.CreateBitCast(Ops[0], FTy);
7561     if (usgn)
7562       return Builder.CreateFPToUI(Ops[0], InTy);
7563     return Builder.CreateFPToSI(Ops[0], InTy);
7564   }
7565   case NEON::BI__builtin_neon_vcvts_f32_u32:
7566   case NEON::BI__builtin_neon_vcvtd_f64_u64:
7567     usgn = true;
7568     LLVM_FALLTHROUGH;
7569   case NEON::BI__builtin_neon_vcvts_f32_s32:
7570   case NEON::BI__builtin_neon_vcvtd_f64_s64: {
7571     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7572     bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
7573     llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
7574     llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
7575     Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
7576     if (usgn)
7577       return Builder.CreateUIToFP(Ops[0], FTy);
7578     return Builder.CreateSIToFP(Ops[0], FTy);
7579   }
7580   case NEON::BI__builtin_neon_vcvth_f16_u16:
7581   case NEON::BI__builtin_neon_vcvth_f16_u32:
7582   case NEON::BI__builtin_neon_vcvth_f16_u64:
7583     usgn = true;
7584     LLVM_FALLTHROUGH;
7585   case NEON::BI__builtin_neon_vcvth_f16_s16:
7586   case NEON::BI__builtin_neon_vcvth_f16_s32:
7587   case NEON::BI__builtin_neon_vcvth_f16_s64: {
7588     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7589     llvm::Type *FTy = HalfTy;
7590     llvm::Type *InTy;
7591     if (Ops[0]->getType()->getPrimitiveSizeInBits() == 64)
7592       InTy = Int64Ty;
7593     else if (Ops[0]->getType()->getPrimitiveSizeInBits() == 32)
7594       InTy = Int32Ty;
7595     else
7596       InTy = Int16Ty;
7597     Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
7598     if (usgn)
7599       return Builder.CreateUIToFP(Ops[0], FTy);
7600     return Builder.CreateSIToFP(Ops[0], FTy);
7601   }
7602   case NEON::BI__builtin_neon_vcvth_u16_f16:
7603     usgn = true;
7604     LLVM_FALLTHROUGH;
7605   case NEON::BI__builtin_neon_vcvth_s16_f16: {
7606     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7607     Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy);
7608     if (usgn)
7609       return Builder.CreateFPToUI(Ops[0], Int16Ty);
7610     return Builder.CreateFPToSI(Ops[0], Int16Ty);
7611   }
7612   case NEON::BI__builtin_neon_vcvth_u32_f16:
7613     usgn = true;
7614     LLVM_FALLTHROUGH;
7615   case NEON::BI__builtin_neon_vcvth_s32_f16: {
7616     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7617     Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy);
7618     if (usgn)
7619       return Builder.CreateFPToUI(Ops[0], Int32Ty);
7620     return Builder.CreateFPToSI(Ops[0], Int32Ty);
7621   }
7622   case NEON::BI__builtin_neon_vcvth_u64_f16:
7623     usgn = true;
7624     LLVM_FALLTHROUGH;
7625   case NEON::BI__builtin_neon_vcvth_s64_f16: {
7626     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7627     Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy);
7628     if (usgn)
7629       return Builder.CreateFPToUI(Ops[0], Int64Ty);
7630     return Builder.CreateFPToSI(Ops[0], Int64Ty);
7631   }
7632   case NEON::BI__builtin_neon_vcvtah_u16_f16:
7633   case NEON::BI__builtin_neon_vcvtmh_u16_f16:
7634   case NEON::BI__builtin_neon_vcvtnh_u16_f16:
7635   case NEON::BI__builtin_neon_vcvtph_u16_f16:
7636   case NEON::BI__builtin_neon_vcvtah_s16_f16:
7637   case NEON::BI__builtin_neon_vcvtmh_s16_f16:
7638   case NEON::BI__builtin_neon_vcvtnh_s16_f16:
7639   case NEON::BI__builtin_neon_vcvtph_s16_f16: {
7640     unsigned Int;
7641     llvm::Type* InTy = Int32Ty;
7642     llvm::Type* FTy  = HalfTy;
7643     llvm::Type *Tys[2] = {InTy, FTy};
7644     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7645     switch (BuiltinID) {
7646     default: llvm_unreachable("missing builtin ID in switch!");
7647     case NEON::BI__builtin_neon_vcvtah_u16_f16:
7648       Int = Intrinsic::aarch64_neon_fcvtau; break;
7649     case NEON::BI__builtin_neon_vcvtmh_u16_f16:
7650       Int = Intrinsic::aarch64_neon_fcvtmu; break;
7651     case NEON::BI__builtin_neon_vcvtnh_u16_f16:
7652       Int = Intrinsic::aarch64_neon_fcvtnu; break;
7653     case NEON::BI__builtin_neon_vcvtph_u16_f16:
7654       Int = Intrinsic::aarch64_neon_fcvtpu; break;
7655     case NEON::BI__builtin_neon_vcvtah_s16_f16:
7656       Int = Intrinsic::aarch64_neon_fcvtas; break;
7657     case NEON::BI__builtin_neon_vcvtmh_s16_f16:
7658       Int = Intrinsic::aarch64_neon_fcvtms; break;
7659     case NEON::BI__builtin_neon_vcvtnh_s16_f16:
7660       Int = Intrinsic::aarch64_neon_fcvtns; break;
7661     case NEON::BI__builtin_neon_vcvtph_s16_f16:
7662       Int = Intrinsic::aarch64_neon_fcvtps; break;
7663     }
7664     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvt");
7665     return Builder.CreateTrunc(Ops[0], Int16Ty);
7666   }
7667   case NEON::BI__builtin_neon_vcaleh_f16:
7668   case NEON::BI__builtin_neon_vcalth_f16:
7669   case NEON::BI__builtin_neon_vcageh_f16:
7670   case NEON::BI__builtin_neon_vcagth_f16: {
7671     unsigned Int;
7672     llvm::Type* InTy = Int32Ty;
7673     llvm::Type* FTy  = HalfTy;
7674     llvm::Type *Tys[2] = {InTy, FTy};
7675     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7676     switch (BuiltinID) {
7677     default: llvm_unreachable("missing builtin ID in switch!");
7678     case NEON::BI__builtin_neon_vcageh_f16:
7679       Int = Intrinsic::aarch64_neon_facge; break;
7680     case NEON::BI__builtin_neon_vcagth_f16:
7681       Int = Intrinsic::aarch64_neon_facgt; break;
7682     case NEON::BI__builtin_neon_vcaleh_f16:
7683       Int = Intrinsic::aarch64_neon_facge; std::swap(Ops[0], Ops[1]); break;
7684     case NEON::BI__builtin_neon_vcalth_f16:
7685       Int = Intrinsic::aarch64_neon_facgt; std::swap(Ops[0], Ops[1]); break;
7686     }
7687     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "facg");
7688     return Builder.CreateTrunc(Ops[0], Int16Ty);
7689   }
7690   case NEON::BI__builtin_neon_vcvth_n_s16_f16:
7691   case NEON::BI__builtin_neon_vcvth_n_u16_f16: {
7692     unsigned Int;
7693     llvm::Type* InTy = Int32Ty;
7694     llvm::Type* FTy  = HalfTy;
7695     llvm::Type *Tys[2] = {InTy, FTy};
7696     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7697     switch (BuiltinID) {
7698     default: llvm_unreachable("missing builtin ID in switch!");
7699     case NEON::BI__builtin_neon_vcvth_n_s16_f16:
7700       Int = Intrinsic::aarch64_neon_vcvtfp2fxs; break;
7701     case NEON::BI__builtin_neon_vcvth_n_u16_f16:
7702       Int = Intrinsic::aarch64_neon_vcvtfp2fxu; break;
7703     }
7704     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n");
7705     return Builder.CreateTrunc(Ops[0], Int16Ty);
7706   }
7707   case NEON::BI__builtin_neon_vcvth_n_f16_s16:
7708   case NEON::BI__builtin_neon_vcvth_n_f16_u16: {
7709     unsigned Int;
7710     llvm::Type* FTy  = HalfTy;
7711     llvm::Type* InTy = Int32Ty;
7712     llvm::Type *Tys[2] = {FTy, InTy};
7713     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7714     switch (BuiltinID) {
7715     default: llvm_unreachable("missing builtin ID in switch!");
7716     case NEON::BI__builtin_neon_vcvth_n_f16_s16:
7717       Int = Intrinsic::aarch64_neon_vcvtfxs2fp;
7718       Ops[0] = Builder.CreateSExt(Ops[0], InTy, "sext");
7719       break;
7720     case NEON::BI__builtin_neon_vcvth_n_f16_u16:
7721       Int = Intrinsic::aarch64_neon_vcvtfxu2fp;
7722       Ops[0] = Builder.CreateZExt(Ops[0], InTy);
7723       break;
7724     }
7725     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n");
7726   }
7727   case NEON::BI__builtin_neon_vpaddd_s64: {
7728     llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2);
7729     Value *Vec = EmitScalarExpr(E->getArg(0));
7730     // The vector is v2f64, so make sure it's bitcast to that.
7731     Vec = Builder.CreateBitCast(Vec, Ty, "v2i64");
7732     llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
7733     llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
7734     Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
7735     Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
7736     // Pairwise addition of a v2f64 into a scalar f64.
7737     return Builder.CreateAdd(Op0, Op1, "vpaddd");
7738   }
7739   case NEON::BI__builtin_neon_vpaddd_f64: {
7740     llvm::Type *Ty =
7741       llvm::VectorType::get(DoubleTy, 2);
7742     Value *Vec = EmitScalarExpr(E->getArg(0));
7743     // The vector is v2f64, so make sure it's bitcast to that.
7744     Vec = Builder.CreateBitCast(Vec, Ty, "v2f64");
7745     llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
7746     llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
7747     Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
7748     Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
7749     // Pairwise addition of a v2f64 into a scalar f64.
7750     return Builder.CreateFAdd(Op0, Op1, "vpaddd");
7751   }
7752   case NEON::BI__builtin_neon_vpadds_f32: {
7753     llvm::Type *Ty =
7754       llvm::VectorType::get(FloatTy, 2);
7755     Value *Vec = EmitScalarExpr(E->getArg(0));
7756     // The vector is v2f32, so make sure it's bitcast to that.
7757     Vec = Builder.CreateBitCast(Vec, Ty, "v2f32");
7758     llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
7759     llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
7760     Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
7761     Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
7762     // Pairwise addition of a v2f32 into a scalar f32.
7763     return Builder.CreateFAdd(Op0, Op1, "vpaddd");
7764   }
7765   case NEON::BI__builtin_neon_vceqzd_s64:
7766   case NEON::BI__builtin_neon_vceqzd_f64:
7767   case NEON::BI__builtin_neon_vceqzs_f32:
7768   case NEON::BI__builtin_neon_vceqzh_f16:
7769     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7770     return EmitAArch64CompareBuiltinExpr(
7771         Ops[0], ConvertType(E->getCallReturnType(getContext())),
7772         ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz");
7773   case NEON::BI__builtin_neon_vcgezd_s64:
7774   case NEON::BI__builtin_neon_vcgezd_f64:
7775   case NEON::BI__builtin_neon_vcgezs_f32:
7776   case NEON::BI__builtin_neon_vcgezh_f16:
7777     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7778     return EmitAArch64CompareBuiltinExpr(
7779         Ops[0], ConvertType(E->getCallReturnType(getContext())),
7780         ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez");
7781   case NEON::BI__builtin_neon_vclezd_s64:
7782   case NEON::BI__builtin_neon_vclezd_f64:
7783   case NEON::BI__builtin_neon_vclezs_f32:
7784   case NEON::BI__builtin_neon_vclezh_f16:
7785     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7786     return EmitAArch64CompareBuiltinExpr(
7787         Ops[0], ConvertType(E->getCallReturnType(getContext())),
7788         ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez");
7789   case NEON::BI__builtin_neon_vcgtzd_s64:
7790   case NEON::BI__builtin_neon_vcgtzd_f64:
7791   case NEON::BI__builtin_neon_vcgtzs_f32:
7792   case NEON::BI__builtin_neon_vcgtzh_f16:
7793     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7794     return EmitAArch64CompareBuiltinExpr(
7795         Ops[0], ConvertType(E->getCallReturnType(getContext())),
7796         ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz");
7797   case NEON::BI__builtin_neon_vcltzd_s64:
7798   case NEON::BI__builtin_neon_vcltzd_f64:
7799   case NEON::BI__builtin_neon_vcltzs_f32:
7800   case NEON::BI__builtin_neon_vcltzh_f16:
7801     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7802     return EmitAArch64CompareBuiltinExpr(
7803         Ops[0], ConvertType(E->getCallReturnType(getContext())),
7804         ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz");
7805 
7806   case NEON::BI__builtin_neon_vceqzd_u64: {
7807     Ops.push_back(EmitScalarExpr(E->getArg(0)));
7808     Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
7809     Ops[0] =
7810         Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty));
7811     return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd");
7812   }
7813   case NEON::BI__builtin_neon_vceqd_f64:
7814   case NEON::BI__builtin_neon_vcled_f64:
7815   case NEON::BI__builtin_neon_vcltd_f64:
7816   case NEON::BI__builtin_neon_vcged_f64:
7817   case NEON::BI__builtin_neon_vcgtd_f64: {
7818     llvm::CmpInst::Predicate P;
7819     switch (BuiltinID) {
7820     default: llvm_unreachable("missing builtin ID in switch!");
7821     case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break;
7822     case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break;
7823     case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break;
7824     case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break;
7825     case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break;
7826     }
7827     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7828     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
7829     Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
7830     Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
7831     return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd");
7832   }
7833   case NEON::BI__builtin_neon_vceqs_f32:
7834   case NEON::BI__builtin_neon_vcles_f32:
7835   case NEON::BI__builtin_neon_vclts_f32:
7836   case NEON::BI__builtin_neon_vcges_f32:
7837   case NEON::BI__builtin_neon_vcgts_f32: {
7838     llvm::CmpInst::Predicate P;
7839     switch (BuiltinID) {
7840     default: llvm_unreachable("missing builtin ID in switch!");
7841     case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break;
7842     case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break;
7843     case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break;
7844     case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break;
7845     case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break;
7846     }
7847     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7848     Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
7849     Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy);
7850     Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
7851     return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd");
7852   }
7853   case NEON::BI__builtin_neon_vceqh_f16:
7854   case NEON::BI__builtin_neon_vcleh_f16:
7855   case NEON::BI__builtin_neon_vclth_f16:
7856   case NEON::BI__builtin_neon_vcgeh_f16:
7857   case NEON::BI__builtin_neon_vcgth_f16: {
7858     llvm::CmpInst::Predicate P;
7859     switch (BuiltinID) {
7860     default: llvm_unreachable("missing builtin ID in switch!");
7861     case NEON::BI__builtin_neon_vceqh_f16: P = llvm::FCmpInst::FCMP_OEQ; break;
7862     case NEON::BI__builtin_neon_vcleh_f16: P = llvm::FCmpInst::FCMP_OLE; break;
7863     case NEON::BI__builtin_neon_vclth_f16: P = llvm::FCmpInst::FCMP_OLT; break;
7864     case NEON::BI__builtin_neon_vcgeh_f16: P = llvm::FCmpInst::FCMP_OGE; break;
7865     case NEON::BI__builtin_neon_vcgth_f16: P = llvm::FCmpInst::FCMP_OGT; break;
7866     }
7867     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7868     Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy);
7869     Ops[1] = Builder.CreateBitCast(Ops[1], HalfTy);
7870     Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
7871     return Builder.CreateSExt(Ops[0], Int16Ty, "vcmpd");
7872   }
7873   case NEON::BI__builtin_neon_vceqd_s64:
7874   case NEON::BI__builtin_neon_vceqd_u64:
7875   case NEON::BI__builtin_neon_vcgtd_s64:
7876   case NEON::BI__builtin_neon_vcgtd_u64:
7877   case NEON::BI__builtin_neon_vcltd_s64:
7878   case NEON::BI__builtin_neon_vcltd_u64:
7879   case NEON::BI__builtin_neon_vcged_u64:
7880   case NEON::BI__builtin_neon_vcged_s64:
7881   case NEON::BI__builtin_neon_vcled_u64:
7882   case NEON::BI__builtin_neon_vcled_s64: {
7883     llvm::CmpInst::Predicate P;
7884     switch (BuiltinID) {
7885     default: llvm_unreachable("missing builtin ID in switch!");
7886     case NEON::BI__builtin_neon_vceqd_s64:
7887     case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break;
7888     case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break;
7889     case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break;
7890     case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break;
7891     case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break;
7892     case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break;
7893     case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break;
7894     case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break;
7895     case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break;
7896     }
7897     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7898     Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
7899     Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
7900     Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]);
7901     return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd");
7902   }
7903   case NEON::BI__builtin_neon_vtstd_s64:
7904   case NEON::BI__builtin_neon_vtstd_u64: {
7905     Ops.push_back(EmitScalarExpr(E->getArg(1)));
7906     Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
7907     Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
7908     Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
7909     Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
7910                                 llvm::Constant::getNullValue(Int64Ty));
7911     return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd");
7912   }
7913   case NEON::BI__builtin_neon_vset_lane_i8:
7914   case NEON::BI__builtin_neon_vset_lane_i16:
7915   case NEON::BI__builtin_neon_vset_lane_i32:
7916   case NEON::BI__builtin_neon_vset_lane_i64:
7917   case NEON::BI__builtin_neon_vset_lane_f32:
7918   case NEON::BI__builtin_neon_vsetq_lane_i8:
7919   case NEON::BI__builtin_neon_vsetq_lane_i16:
7920   case NEON::BI__builtin_neon_vsetq_lane_i32:
7921   case NEON::BI__builtin_neon_vsetq_lane_i64:
7922   case NEON::BI__builtin_neon_vsetq_lane_f32:
7923     Ops.push_back(EmitScalarExpr(E->getArg(2)));
7924     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
7925   case NEON::BI__builtin_neon_vset_lane_f64:
7926     // The vector type needs a cast for the v1f64 variant.
7927     Ops[1] = Builder.CreateBitCast(Ops[1],
7928                                    llvm::VectorType::get(DoubleTy, 1));
7929     Ops.push_back(EmitScalarExpr(E->getArg(2)));
7930     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
7931   case NEON::BI__builtin_neon_vsetq_lane_f64:
7932     // The vector type needs a cast for the v2f64 variant.
7933     Ops[1] = Builder.CreateBitCast(Ops[1],
7934         llvm::VectorType::get(DoubleTy, 2));
7935     Ops.push_back(EmitScalarExpr(E->getArg(2)));
7936     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
7937 
7938   case NEON::BI__builtin_neon_vget_lane_i8:
7939   case NEON::BI__builtin_neon_vdupb_lane_i8:
7940     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8));
7941     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7942                                         "vget_lane");
7943   case NEON::BI__builtin_neon_vgetq_lane_i8:
7944   case NEON::BI__builtin_neon_vdupb_laneq_i8:
7945     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16));
7946     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7947                                         "vgetq_lane");
7948   case NEON::BI__builtin_neon_vget_lane_i16:
7949   case NEON::BI__builtin_neon_vduph_lane_i16:
7950     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4));
7951     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7952                                         "vget_lane");
7953   case NEON::BI__builtin_neon_vgetq_lane_i16:
7954   case NEON::BI__builtin_neon_vduph_laneq_i16:
7955     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8));
7956     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7957                                         "vgetq_lane");
7958   case NEON::BI__builtin_neon_vget_lane_i32:
7959   case NEON::BI__builtin_neon_vdups_lane_i32:
7960     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2));
7961     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7962                                         "vget_lane");
7963   case NEON::BI__builtin_neon_vdups_lane_f32:
7964     Ops[0] = Builder.CreateBitCast(Ops[0],
7965         llvm::VectorType::get(FloatTy, 2));
7966     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7967                                         "vdups_lane");
7968   case NEON::BI__builtin_neon_vgetq_lane_i32:
7969   case NEON::BI__builtin_neon_vdups_laneq_i32:
7970     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4));
7971     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7972                                         "vgetq_lane");
7973   case NEON::BI__builtin_neon_vget_lane_i64:
7974   case NEON::BI__builtin_neon_vdupd_lane_i64:
7975     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1));
7976     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7977                                         "vget_lane");
7978   case NEON::BI__builtin_neon_vdupd_lane_f64:
7979     Ops[0] = Builder.CreateBitCast(Ops[0],
7980         llvm::VectorType::get(DoubleTy, 1));
7981     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7982                                         "vdupd_lane");
7983   case NEON::BI__builtin_neon_vgetq_lane_i64:
7984   case NEON::BI__builtin_neon_vdupd_laneq_i64:
7985     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2));
7986     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7987                                         "vgetq_lane");
7988   case NEON::BI__builtin_neon_vget_lane_f32:
7989     Ops[0] = Builder.CreateBitCast(Ops[0],
7990         llvm::VectorType::get(FloatTy, 2));
7991     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7992                                         "vget_lane");
7993   case NEON::BI__builtin_neon_vget_lane_f64:
7994     Ops[0] = Builder.CreateBitCast(Ops[0],
7995         llvm::VectorType::get(DoubleTy, 1));
7996     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
7997                                         "vget_lane");
7998   case NEON::BI__builtin_neon_vgetq_lane_f32:
7999   case NEON::BI__builtin_neon_vdups_laneq_f32:
8000     Ops[0] = Builder.CreateBitCast(Ops[0],
8001         llvm::VectorType::get(FloatTy, 4));
8002     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
8003                                         "vgetq_lane");
8004   case NEON::BI__builtin_neon_vgetq_lane_f64:
8005   case NEON::BI__builtin_neon_vdupd_laneq_f64:
8006     Ops[0] = Builder.CreateBitCast(Ops[0],
8007         llvm::VectorType::get(DoubleTy, 2));
8008     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
8009                                         "vgetq_lane");
8010   case NEON::BI__builtin_neon_vaddh_f16:
8011     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8012     return Builder.CreateFAdd(Ops[0], Ops[1], "vaddh");
8013   case NEON::BI__builtin_neon_vsubh_f16:
8014     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8015     return Builder.CreateFSub(Ops[0], Ops[1], "vsubh");
8016   case NEON::BI__builtin_neon_vmulh_f16:
8017     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8018     return Builder.CreateFMul(Ops[0], Ops[1], "vmulh");
8019   case NEON::BI__builtin_neon_vdivh_f16:
8020     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8021     return Builder.CreateFDiv(Ops[0], Ops[1], "vdivh");
8022   case NEON::BI__builtin_neon_vfmah_f16: {
8023     Function *F = CGM.getIntrinsic(Intrinsic::fma, HalfTy);
8024     // NEON intrinsic puts accumulator first, unlike the LLVM fma.
8025     return Builder.CreateCall(F,
8026       {EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), Ops[0]});
8027   }
8028   case NEON::BI__builtin_neon_vfmsh_f16: {
8029     Function *F = CGM.getIntrinsic(Intrinsic::fma, HalfTy);
8030     Value *Zero = llvm::ConstantFP::getZeroValueForNegation(HalfTy);
8031     Value* Sub = Builder.CreateFSub(Zero, EmitScalarExpr(E->getArg(1)), "vsubh");
8032     // NEON intrinsic puts accumulator first, unlike the LLVM fma.
8033     return Builder.CreateCall(F, {Sub, EmitScalarExpr(E->getArg(2)), Ops[0]});
8034   }
8035   case NEON::BI__builtin_neon_vaddd_s64:
8036   case NEON::BI__builtin_neon_vaddd_u64:
8037     return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd");
8038   case NEON::BI__builtin_neon_vsubd_s64:
8039   case NEON::BI__builtin_neon_vsubd_u64:
8040     return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd");
8041   case NEON::BI__builtin_neon_vqdmlalh_s16:
8042   case NEON::BI__builtin_neon_vqdmlslh_s16: {
8043     SmallVector<Value *, 2> ProductOps;
8044     ProductOps.push_back(vectorWrapScalar16(Ops[1]));
8045     ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2))));
8046     llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
8047     Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
8048                           ProductOps, "vqdmlXl");
8049     Constant *CI = ConstantInt::get(SizeTy, 0);
8050     Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
8051 
8052     unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16
8053                                         ? Intrinsic::aarch64_neon_sqadd
8054                                         : Intrinsic::aarch64_neon_sqsub;
8055     return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl");
8056   }
8057   case NEON::BI__builtin_neon_vqshlud_n_s64: {
8058     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8059     Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
8060     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty),
8061                         Ops, "vqshlu_n");
8062   }
8063   case NEON::BI__builtin_neon_vqshld_n_u64:
8064   case NEON::BI__builtin_neon_vqshld_n_s64: {
8065     unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64
8066                                    ? Intrinsic::aarch64_neon_uqshl
8067                                    : Intrinsic::aarch64_neon_sqshl;
8068     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8069     Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
8070     return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n");
8071   }
8072   case NEON::BI__builtin_neon_vrshrd_n_u64:
8073   case NEON::BI__builtin_neon_vrshrd_n_s64: {
8074     unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64
8075                                    ? Intrinsic::aarch64_neon_urshl
8076                                    : Intrinsic::aarch64_neon_srshl;
8077     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8078     int SV = cast<ConstantInt>(Ops[1])->getSExtValue();
8079     Ops[1] = ConstantInt::get(Int64Ty, -SV);
8080     return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n");
8081   }
8082   case NEON::BI__builtin_neon_vrsrad_n_u64:
8083   case NEON::BI__builtin_neon_vrsrad_n_s64: {
8084     unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64
8085                                    ? Intrinsic::aarch64_neon_urshl
8086                                    : Intrinsic::aarch64_neon_srshl;
8087     Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
8088     Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2))));
8089     Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty),
8090                                 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)});
8091     return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty));
8092   }
8093   case NEON::BI__builtin_neon_vshld_n_s64:
8094   case NEON::BI__builtin_neon_vshld_n_u64: {
8095     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
8096     return Builder.CreateShl(
8097         Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n");
8098   }
8099   case NEON::BI__builtin_neon_vshrd_n_s64: {
8100     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
8101     return Builder.CreateAShr(
8102         Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
8103                                                    Amt->getZExtValue())),
8104         "shrd_n");
8105   }
8106   case NEON::BI__builtin_neon_vshrd_n_u64: {
8107     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
8108     uint64_t ShiftAmt = Amt->getZExtValue();
8109     // Right-shifting an unsigned value by its size yields 0.
8110     if (ShiftAmt == 64)
8111       return ConstantInt::get(Int64Ty, 0);
8112     return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt),
8113                               "shrd_n");
8114   }
8115   case NEON::BI__builtin_neon_vsrad_n_s64: {
8116     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
8117     Ops[1] = Builder.CreateAShr(
8118         Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
8119                                                    Amt->getZExtValue())),
8120         "shrd_n");
8121     return Builder.CreateAdd(Ops[0], Ops[1]);
8122   }
8123   case NEON::BI__builtin_neon_vsrad_n_u64: {
8124     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
8125     uint64_t ShiftAmt = Amt->getZExtValue();
8126     // Right-shifting an unsigned value by its size yields 0.
8127     // As Op + 0 = Op, return Ops[0] directly.
8128     if (ShiftAmt == 64)
8129       return Ops[0];
8130     Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt),
8131                                 "shrd_n");
8132     return Builder.CreateAdd(Ops[0], Ops[1]);
8133   }
8134   case NEON::BI__builtin_neon_vqdmlalh_lane_s16:
8135   case NEON::BI__builtin_neon_vqdmlalh_laneq_s16:
8136   case NEON::BI__builtin_neon_vqdmlslh_lane_s16:
8137   case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: {
8138     Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
8139                                           "lane");
8140     SmallVector<Value *, 2> ProductOps;
8141     ProductOps.push_back(vectorWrapScalar16(Ops[1]));
8142     ProductOps.push_back(vectorWrapScalar16(Ops[2]));
8143     llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
8144     Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
8145                           ProductOps, "vqdmlXl");
8146     Constant *CI = ConstantInt::get(SizeTy, 0);
8147     Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
8148     Ops.pop_back();
8149 
8150     unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 ||
8151                        BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16)
8152                           ? Intrinsic::aarch64_neon_sqadd
8153                           : Intrinsic::aarch64_neon_sqsub;
8154     return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl");
8155   }
8156   case NEON::BI__builtin_neon_vqdmlals_s32:
8157   case NEON::BI__builtin_neon_vqdmlsls_s32: {
8158     SmallVector<Value *, 2> ProductOps;
8159     ProductOps.push_back(Ops[1]);
8160     ProductOps.push_back(EmitScalarExpr(E->getArg(2)));
8161     Ops[1] =
8162         EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
8163                      ProductOps, "vqdmlXl");
8164 
8165     unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32
8166                                         ? Intrinsic::aarch64_neon_sqadd
8167                                         : Intrinsic::aarch64_neon_sqsub;
8168     return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl");
8169   }
8170   case NEON::BI__builtin_neon_vqdmlals_lane_s32:
8171   case NEON::BI__builtin_neon_vqdmlals_laneq_s32:
8172   case NEON::BI__builtin_neon_vqdmlsls_lane_s32:
8173   case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: {
8174     Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
8175                                           "lane");
8176     SmallVector<Value *, 2> ProductOps;
8177     ProductOps.push_back(Ops[1]);
8178     ProductOps.push_back(Ops[2]);
8179     Ops[1] =
8180         EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
8181                      ProductOps, "vqdmlXl");
8182     Ops.pop_back();
8183 
8184     unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 ||
8185                        BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32)
8186                           ? Intrinsic::aarch64_neon_sqadd
8187                           : Intrinsic::aarch64_neon_sqsub;
8188     return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl");
8189   }
8190   case NEON::BI__builtin_neon_vduph_lane_f16: {
8191     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
8192                                         "vget_lane");
8193   }
8194   case NEON::BI__builtin_neon_vduph_laneq_f16: {
8195     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
8196                                         "vgetq_lane");
8197   }
8198   case AArch64::BI_BitScanForward:
8199   case AArch64::BI_BitScanForward64:
8200     return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E);
8201   case AArch64::BI_BitScanReverse:
8202   case AArch64::BI_BitScanReverse64:
8203     return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E);
8204   case AArch64::BI_InterlockedAnd64:
8205     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E);
8206   case AArch64::BI_InterlockedExchange64:
8207     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E);
8208   case AArch64::BI_InterlockedExchangeAdd64:
8209     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E);
8210   case AArch64::BI_InterlockedExchangeSub64:
8211     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E);
8212   case AArch64::BI_InterlockedOr64:
8213     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E);
8214   case AArch64::BI_InterlockedXor64:
8215     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E);
8216   case AArch64::BI_InterlockedDecrement64:
8217     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E);
8218   case AArch64::BI_InterlockedIncrement64:
8219     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E);
8220   case AArch64::BI_InterlockedExchangeAdd8_acq:
8221   case AArch64::BI_InterlockedExchangeAdd16_acq:
8222   case AArch64::BI_InterlockedExchangeAdd_acq:
8223   case AArch64::BI_InterlockedExchangeAdd64_acq:
8224     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_acq, E);
8225   case AArch64::BI_InterlockedExchangeAdd8_rel:
8226   case AArch64::BI_InterlockedExchangeAdd16_rel:
8227   case AArch64::BI_InterlockedExchangeAdd_rel:
8228   case AArch64::BI_InterlockedExchangeAdd64_rel:
8229     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_rel, E);
8230   case AArch64::BI_InterlockedExchangeAdd8_nf:
8231   case AArch64::BI_InterlockedExchangeAdd16_nf:
8232   case AArch64::BI_InterlockedExchangeAdd_nf:
8233   case AArch64::BI_InterlockedExchangeAdd64_nf:
8234     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_nf, E);
8235   case AArch64::BI_InterlockedExchange8_acq:
8236   case AArch64::BI_InterlockedExchange16_acq:
8237   case AArch64::BI_InterlockedExchange_acq:
8238   case AArch64::BI_InterlockedExchange64_acq:
8239     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_acq, E);
8240   case AArch64::BI_InterlockedExchange8_rel:
8241   case AArch64::BI_InterlockedExchange16_rel:
8242   case AArch64::BI_InterlockedExchange_rel:
8243   case AArch64::BI_InterlockedExchange64_rel:
8244     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_rel, E);
8245   case AArch64::BI_InterlockedExchange8_nf:
8246   case AArch64::BI_InterlockedExchange16_nf:
8247   case AArch64::BI_InterlockedExchange_nf:
8248   case AArch64::BI_InterlockedExchange64_nf:
8249     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_nf, E);
8250   case AArch64::BI_InterlockedCompareExchange8_acq:
8251   case AArch64::BI_InterlockedCompareExchange16_acq:
8252   case AArch64::BI_InterlockedCompareExchange_acq:
8253   case AArch64::BI_InterlockedCompareExchange64_acq:
8254     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_acq, E);
8255   case AArch64::BI_InterlockedCompareExchange8_rel:
8256   case AArch64::BI_InterlockedCompareExchange16_rel:
8257   case AArch64::BI_InterlockedCompareExchange_rel:
8258   case AArch64::BI_InterlockedCompareExchange64_rel:
8259     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_rel, E);
8260   case AArch64::BI_InterlockedCompareExchange8_nf:
8261   case AArch64::BI_InterlockedCompareExchange16_nf:
8262   case AArch64::BI_InterlockedCompareExchange_nf:
8263   case AArch64::BI_InterlockedCompareExchange64_nf:
8264     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_nf, E);
8265   case AArch64::BI_InterlockedOr8_acq:
8266   case AArch64::BI_InterlockedOr16_acq:
8267   case AArch64::BI_InterlockedOr_acq:
8268   case AArch64::BI_InterlockedOr64_acq:
8269     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_acq, E);
8270   case AArch64::BI_InterlockedOr8_rel:
8271   case AArch64::BI_InterlockedOr16_rel:
8272   case AArch64::BI_InterlockedOr_rel:
8273   case AArch64::BI_InterlockedOr64_rel:
8274     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_rel, E);
8275   case AArch64::BI_InterlockedOr8_nf:
8276   case AArch64::BI_InterlockedOr16_nf:
8277   case AArch64::BI_InterlockedOr_nf:
8278   case AArch64::BI_InterlockedOr64_nf:
8279     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_nf, E);
8280   case AArch64::BI_InterlockedXor8_acq:
8281   case AArch64::BI_InterlockedXor16_acq:
8282   case AArch64::BI_InterlockedXor_acq:
8283   case AArch64::BI_InterlockedXor64_acq:
8284     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_acq, E);
8285   case AArch64::BI_InterlockedXor8_rel:
8286   case AArch64::BI_InterlockedXor16_rel:
8287   case AArch64::BI_InterlockedXor_rel:
8288   case AArch64::BI_InterlockedXor64_rel:
8289     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_rel, E);
8290   case AArch64::BI_InterlockedXor8_nf:
8291   case AArch64::BI_InterlockedXor16_nf:
8292   case AArch64::BI_InterlockedXor_nf:
8293   case AArch64::BI_InterlockedXor64_nf:
8294     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_nf, E);
8295   case AArch64::BI_InterlockedAnd8_acq:
8296   case AArch64::BI_InterlockedAnd16_acq:
8297   case AArch64::BI_InterlockedAnd_acq:
8298   case AArch64::BI_InterlockedAnd64_acq:
8299     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_acq, E);
8300   case AArch64::BI_InterlockedAnd8_rel:
8301   case AArch64::BI_InterlockedAnd16_rel:
8302   case AArch64::BI_InterlockedAnd_rel:
8303   case AArch64::BI_InterlockedAnd64_rel:
8304     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_rel, E);
8305   case AArch64::BI_InterlockedAnd8_nf:
8306   case AArch64::BI_InterlockedAnd16_nf:
8307   case AArch64::BI_InterlockedAnd_nf:
8308   case AArch64::BI_InterlockedAnd64_nf:
8309     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_nf, E);
8310   case AArch64::BI_InterlockedIncrement16_acq:
8311   case AArch64::BI_InterlockedIncrement_acq:
8312   case AArch64::BI_InterlockedIncrement64_acq:
8313     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_acq, E);
8314   case AArch64::BI_InterlockedIncrement16_rel:
8315   case AArch64::BI_InterlockedIncrement_rel:
8316   case AArch64::BI_InterlockedIncrement64_rel:
8317     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_rel, E);
8318   case AArch64::BI_InterlockedIncrement16_nf:
8319   case AArch64::BI_InterlockedIncrement_nf:
8320   case AArch64::BI_InterlockedIncrement64_nf:
8321     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_nf, E);
8322   case AArch64::BI_InterlockedDecrement16_acq:
8323   case AArch64::BI_InterlockedDecrement_acq:
8324   case AArch64::BI_InterlockedDecrement64_acq:
8325     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_acq, E);
8326   case AArch64::BI_InterlockedDecrement16_rel:
8327   case AArch64::BI_InterlockedDecrement_rel:
8328   case AArch64::BI_InterlockedDecrement64_rel:
8329     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_rel, E);
8330   case AArch64::BI_InterlockedDecrement16_nf:
8331   case AArch64::BI_InterlockedDecrement_nf:
8332   case AArch64::BI_InterlockedDecrement64_nf:
8333     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_nf, E);
8334 
8335   case AArch64::BI_InterlockedAdd: {
8336     Value *Arg0 = EmitScalarExpr(E->getArg(0));
8337     Value *Arg1 = EmitScalarExpr(E->getArg(1));
8338     AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
8339       AtomicRMWInst::Add, Arg0, Arg1,
8340       llvm::AtomicOrdering::SequentiallyConsistent);
8341     return Builder.CreateAdd(RMWI, Arg1);
8342   }
8343   }
8344 
8345   llvm::VectorType *VTy = GetNeonType(this, Type);
8346   llvm::Type *Ty = VTy;
8347   if (!Ty)
8348     return nullptr;
8349 
8350   // Not all intrinsics handled by the common case work for AArch64 yet, so only
8351   // defer to common code if it's been added to our special map.
8352   Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID,
8353                                    AArch64SIMDIntrinsicsProvenSorted);
8354 
8355   if (Builtin)
8356     return EmitCommonNeonBuiltinExpr(
8357         Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
8358         Builtin->NameHint, Builtin->TypeModifier, E, Ops,
8359         /*never use addresses*/ Address::invalid(), Address::invalid(), Arch);
8360 
8361   if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops, Arch))
8362     return V;
8363 
8364   unsigned Int;
8365   switch (BuiltinID) {
8366   default: return nullptr;
8367   case NEON::BI__builtin_neon_vbsl_v:
8368   case NEON::BI__builtin_neon_vbslq_v: {
8369     llvm::Type *BitTy = llvm::VectorType::getInteger(VTy);
8370     Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl");
8371     Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl");
8372     Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl");
8373 
8374     Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl");
8375     Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl");
8376     Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl");
8377     return Builder.CreateBitCast(Ops[0], Ty);
8378   }
8379   case NEON::BI__builtin_neon_vfma_lane_v:
8380   case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types
8381     // The ARM builtins (and instructions) have the addend as the first
8382     // operand, but the 'fma' intrinsics have it last. Swap it around here.
8383     Value *Addend = Ops[0];
8384     Value *Multiplicand = Ops[1];
8385     Value *LaneSource = Ops[2];
8386     Ops[0] = Multiplicand;
8387     Ops[1] = LaneSource;
8388     Ops[2] = Addend;
8389 
8390     // Now adjust things to handle the lane access.
8391     llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ?
8392       llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) :
8393       VTy;
8394     llvm::Constant *cst = cast<Constant>(Ops[3]);
8395     Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst);
8396     Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy);
8397     Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane");
8398 
8399     Ops.pop_back();
8400     Int = Intrinsic::fma;
8401     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla");
8402   }
8403   case NEON::BI__builtin_neon_vfma_laneq_v: {
8404     llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
8405     // v1f64 fma should be mapped to Neon scalar f64 fma
8406     if (VTy && VTy->getElementType() == DoubleTy) {
8407       Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
8408       Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
8409       llvm::Type *VTy = GetNeonType(this,
8410         NeonTypeFlags(NeonTypeFlags::Float64, false, true));
8411       Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
8412       Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
8413       Function *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy);
8414       Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]});
8415       return Builder.CreateBitCast(Result, Ty);
8416     }
8417     Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
8418     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
8419     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8420 
8421     llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(),
8422                                             VTy->getNumElements() * 2);
8423     Ops[2] = Builder.CreateBitCast(Ops[2], STy);
8424     Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(),
8425                                                cast<ConstantInt>(Ops[3]));
8426     Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane");
8427 
8428     return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]});
8429   }
8430   case NEON::BI__builtin_neon_vfmaq_laneq_v: {
8431     Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
8432     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
8433     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8434 
8435     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
8436     Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3]));
8437     return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]});
8438   }
8439   case NEON::BI__builtin_neon_vfmah_lane_f16:
8440   case NEON::BI__builtin_neon_vfmas_lane_f32:
8441   case NEON::BI__builtin_neon_vfmah_laneq_f16:
8442   case NEON::BI__builtin_neon_vfmas_laneq_f32:
8443   case NEON::BI__builtin_neon_vfmad_lane_f64:
8444   case NEON::BI__builtin_neon_vfmad_laneq_f64: {
8445     Ops.push_back(EmitScalarExpr(E->getArg(3)));
8446     llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
8447     Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
8448     Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
8449     return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]});
8450   }
8451   case NEON::BI__builtin_neon_vmull_v:
8452     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
8453     Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull;
8454     if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull;
8455     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
8456   case NEON::BI__builtin_neon_vmax_v:
8457   case NEON::BI__builtin_neon_vmaxq_v:
8458     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
8459     Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax;
8460     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax;
8461     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
8462   case NEON::BI__builtin_neon_vmaxh_f16: {
8463     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8464     Int = Intrinsic::aarch64_neon_fmax;
8465     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmax");
8466   }
8467   case NEON::BI__builtin_neon_vmin_v:
8468   case NEON::BI__builtin_neon_vminq_v:
8469     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
8470     Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin;
8471     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin;
8472     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
8473   case NEON::BI__builtin_neon_vminh_f16: {
8474     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8475     Int = Intrinsic::aarch64_neon_fmin;
8476     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmin");
8477   }
8478   case NEON::BI__builtin_neon_vabd_v:
8479   case NEON::BI__builtin_neon_vabdq_v:
8480     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
8481     Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd;
8482     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd;
8483     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
8484   case NEON::BI__builtin_neon_vpadal_v:
8485   case NEON::BI__builtin_neon_vpadalq_v: {
8486     unsigned ArgElts = VTy->getNumElements();
8487     llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType());
8488     unsigned BitWidth = EltTy->getBitWidth();
8489     llvm::Type *ArgTy = llvm::VectorType::get(
8490         llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts);
8491     llvm::Type* Tys[2] = { VTy, ArgTy };
8492     Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp;
8493     SmallVector<llvm::Value*, 1> TmpOps;
8494     TmpOps.push_back(Ops[1]);
8495     Function *F = CGM.getIntrinsic(Int, Tys);
8496     llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal");
8497     llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType());
8498     return Builder.CreateAdd(tmp, addend);
8499   }
8500   case NEON::BI__builtin_neon_vpmin_v:
8501   case NEON::BI__builtin_neon_vpminq_v:
8502     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
8503     Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp;
8504     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp;
8505     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
8506   case NEON::BI__builtin_neon_vpmax_v:
8507   case NEON::BI__builtin_neon_vpmaxq_v:
8508     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
8509     Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp;
8510     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp;
8511     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
8512   case NEON::BI__builtin_neon_vminnm_v:
8513   case NEON::BI__builtin_neon_vminnmq_v:
8514     Int = Intrinsic::aarch64_neon_fminnm;
8515     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm");
8516   case NEON::BI__builtin_neon_vminnmh_f16:
8517     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8518     Int = Intrinsic::aarch64_neon_fminnm;
8519     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vminnm");
8520   case NEON::BI__builtin_neon_vmaxnm_v:
8521   case NEON::BI__builtin_neon_vmaxnmq_v:
8522     Int = Intrinsic::aarch64_neon_fmaxnm;
8523     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm");
8524   case NEON::BI__builtin_neon_vmaxnmh_f16:
8525     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8526     Int = Intrinsic::aarch64_neon_fmaxnm;
8527     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmaxnm");
8528   case NEON::BI__builtin_neon_vrecpss_f32: {
8529     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8530     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy),
8531                         Ops, "vrecps");
8532   }
8533   case NEON::BI__builtin_neon_vrecpsd_f64:
8534     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8535     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy),
8536                         Ops, "vrecps");
8537   case NEON::BI__builtin_neon_vrecpsh_f16:
8538     Ops.push_back(EmitScalarExpr(E->getArg(1)));
8539     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, HalfTy),
8540                         Ops, "vrecps");
8541   case NEON::BI__builtin_neon_vqshrun_n_v:
8542     Int = Intrinsic::aarch64_neon_sqshrun;
8543     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n");
8544   case NEON::BI__builtin_neon_vqrshrun_n_v:
8545     Int = Intrinsic::aarch64_neon_sqrshrun;
8546     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n");
8547   case NEON::BI__builtin_neon_vqshrn_n_v:
8548     Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn;
8549     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n");
8550   case NEON::BI__builtin_neon_vrshrn_n_v:
8551     Int = Intrinsic::aarch64_neon_rshrn;
8552     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n");
8553   case NEON::BI__builtin_neon_vqrshrn_n_v:
8554     Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn;
8555     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n");
8556   case NEON::BI__builtin_neon_vrndah_f16: {
8557     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8558     Int = Intrinsic::round;
8559     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrnda");
8560   }
8561   case NEON::BI__builtin_neon_vrnda_v:
8562   case NEON::BI__builtin_neon_vrndaq_v: {
8563     Int = Intrinsic::round;
8564     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda");
8565   }
8566   case NEON::BI__builtin_neon_vrndih_f16: {
8567     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8568     Int = Intrinsic::nearbyint;
8569     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndi");
8570   }
8571   case NEON::BI__builtin_neon_vrndmh_f16: {
8572     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8573     Int = Intrinsic::floor;
8574     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndm");
8575   }
8576   case NEON::BI__builtin_neon_vrndm_v:
8577   case NEON::BI__builtin_neon_vrndmq_v: {
8578     Int = Intrinsic::floor;
8579     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm");
8580   }
8581   case NEON::BI__builtin_neon_vrndnh_f16: {
8582     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8583     Int = Intrinsic::aarch64_neon_frintn;
8584     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndn");
8585   }
8586   case NEON::BI__builtin_neon_vrndn_v:
8587   case NEON::BI__builtin_neon_vrndnq_v: {
8588     Int = Intrinsic::aarch64_neon_frintn;
8589     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn");
8590   }
8591   case NEON::BI__builtin_neon_vrndns_f32: {
8592     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8593     Int = Intrinsic::aarch64_neon_frintn;
8594     return EmitNeonCall(CGM.getIntrinsic(Int, FloatTy), Ops, "vrndn");
8595   }
8596   case NEON::BI__builtin_neon_vrndph_f16: {
8597     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8598     Int = Intrinsic::ceil;
8599     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndp");
8600   }
8601   case NEON::BI__builtin_neon_vrndp_v:
8602   case NEON::BI__builtin_neon_vrndpq_v: {
8603     Int = Intrinsic::ceil;
8604     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp");
8605   }
8606   case NEON::BI__builtin_neon_vrndxh_f16: {
8607     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8608     Int = Intrinsic::rint;
8609     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndx");
8610   }
8611   case NEON::BI__builtin_neon_vrndx_v:
8612   case NEON::BI__builtin_neon_vrndxq_v: {
8613     Int = Intrinsic::rint;
8614     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx");
8615   }
8616   case NEON::BI__builtin_neon_vrndh_f16: {
8617     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8618     Int = Intrinsic::trunc;
8619     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndz");
8620   }
8621   case NEON::BI__builtin_neon_vrnd_v:
8622   case NEON::BI__builtin_neon_vrndq_v: {
8623     Int = Intrinsic::trunc;
8624     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz");
8625   }
8626   case NEON::BI__builtin_neon_vcvt_f64_v:
8627   case NEON::BI__builtin_neon_vcvtq_f64_v:
8628     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
8629     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad));
8630     return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
8631                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
8632   case NEON::BI__builtin_neon_vcvt_f64_f32: {
8633     assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&
8634            "unexpected vcvt_f64_f32 builtin");
8635     NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false);
8636     Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
8637 
8638     return Builder.CreateFPExt(Ops[0], Ty, "vcvt");
8639   }
8640   case NEON::BI__builtin_neon_vcvt_f32_f64: {
8641     assert(Type.getEltType() == NeonTypeFlags::Float32 &&
8642            "unexpected vcvt_f32_f64 builtin");
8643     NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true);
8644     Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
8645 
8646     return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt");
8647   }
8648   case NEON::BI__builtin_neon_vcvt_s32_v:
8649   case NEON::BI__builtin_neon_vcvt_u32_v:
8650   case NEON::BI__builtin_neon_vcvt_s64_v:
8651   case NEON::BI__builtin_neon_vcvt_u64_v:
8652   case NEON::BI__builtin_neon_vcvt_s16_v:
8653   case NEON::BI__builtin_neon_vcvt_u16_v:
8654   case NEON::BI__builtin_neon_vcvtq_s32_v:
8655   case NEON::BI__builtin_neon_vcvtq_u32_v:
8656   case NEON::BI__builtin_neon_vcvtq_s64_v:
8657   case NEON::BI__builtin_neon_vcvtq_u64_v:
8658   case NEON::BI__builtin_neon_vcvtq_s16_v:
8659   case NEON::BI__builtin_neon_vcvtq_u16_v: {
8660     Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type));
8661     if (usgn)
8662       return Builder.CreateFPToUI(Ops[0], Ty);
8663     return Builder.CreateFPToSI(Ops[0], Ty);
8664   }
8665   case NEON::BI__builtin_neon_vcvta_s16_v:
8666   case NEON::BI__builtin_neon_vcvta_u16_v:
8667   case NEON::BI__builtin_neon_vcvta_s32_v:
8668   case NEON::BI__builtin_neon_vcvtaq_s16_v:
8669   case NEON::BI__builtin_neon_vcvtaq_s32_v:
8670   case NEON::BI__builtin_neon_vcvta_u32_v:
8671   case NEON::BI__builtin_neon_vcvtaq_u16_v:
8672   case NEON::BI__builtin_neon_vcvtaq_u32_v:
8673   case NEON::BI__builtin_neon_vcvta_s64_v:
8674   case NEON::BI__builtin_neon_vcvtaq_s64_v:
8675   case NEON::BI__builtin_neon_vcvta_u64_v:
8676   case NEON::BI__builtin_neon_vcvtaq_u64_v: {
8677     Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas;
8678     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
8679     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta");
8680   }
8681   case NEON::BI__builtin_neon_vcvtm_s16_v:
8682   case NEON::BI__builtin_neon_vcvtm_s32_v:
8683   case NEON::BI__builtin_neon_vcvtmq_s16_v:
8684   case NEON::BI__builtin_neon_vcvtmq_s32_v:
8685   case NEON::BI__builtin_neon_vcvtm_u16_v:
8686   case NEON::BI__builtin_neon_vcvtm_u32_v:
8687   case NEON::BI__builtin_neon_vcvtmq_u16_v:
8688   case NEON::BI__builtin_neon_vcvtmq_u32_v:
8689   case NEON::BI__builtin_neon_vcvtm_s64_v:
8690   case NEON::BI__builtin_neon_vcvtmq_s64_v:
8691   case NEON::BI__builtin_neon_vcvtm_u64_v:
8692   case NEON::BI__builtin_neon_vcvtmq_u64_v: {
8693     Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms;
8694     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
8695     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm");
8696   }
8697   case NEON::BI__builtin_neon_vcvtn_s16_v:
8698   case NEON::BI__builtin_neon_vcvtn_s32_v:
8699   case NEON::BI__builtin_neon_vcvtnq_s16_v:
8700   case NEON::BI__builtin_neon_vcvtnq_s32_v:
8701   case NEON::BI__builtin_neon_vcvtn_u16_v:
8702   case NEON::BI__builtin_neon_vcvtn_u32_v:
8703   case NEON::BI__builtin_neon_vcvtnq_u16_v:
8704   case NEON::BI__builtin_neon_vcvtnq_u32_v:
8705   case NEON::BI__builtin_neon_vcvtn_s64_v:
8706   case NEON::BI__builtin_neon_vcvtnq_s64_v:
8707   case NEON::BI__builtin_neon_vcvtn_u64_v:
8708   case NEON::BI__builtin_neon_vcvtnq_u64_v: {
8709     Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns;
8710     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
8711     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn");
8712   }
8713   case NEON::BI__builtin_neon_vcvtp_s16_v:
8714   case NEON::BI__builtin_neon_vcvtp_s32_v:
8715   case NEON::BI__builtin_neon_vcvtpq_s16_v:
8716   case NEON::BI__builtin_neon_vcvtpq_s32_v:
8717   case NEON::BI__builtin_neon_vcvtp_u16_v:
8718   case NEON::BI__builtin_neon_vcvtp_u32_v:
8719   case NEON::BI__builtin_neon_vcvtpq_u16_v:
8720   case NEON::BI__builtin_neon_vcvtpq_u32_v:
8721   case NEON::BI__builtin_neon_vcvtp_s64_v:
8722   case NEON::BI__builtin_neon_vcvtpq_s64_v:
8723   case NEON::BI__builtin_neon_vcvtp_u64_v:
8724   case NEON::BI__builtin_neon_vcvtpq_u64_v: {
8725     Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps;
8726     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
8727     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp");
8728   }
8729   case NEON::BI__builtin_neon_vmulx_v:
8730   case NEON::BI__builtin_neon_vmulxq_v: {
8731     Int = Intrinsic::aarch64_neon_fmulx;
8732     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
8733   }
8734   case NEON::BI__builtin_neon_vmulxh_lane_f16:
8735   case NEON::BI__builtin_neon_vmulxh_laneq_f16: {
8736     // vmulx_lane should be mapped to Neon scalar mulx after
8737     // extracting the scalar element
8738     Ops.push_back(EmitScalarExpr(E->getArg(2)));
8739     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
8740     Ops.pop_back();
8741     Int = Intrinsic::aarch64_neon_fmulx;
8742     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmulx");
8743   }
8744   case NEON::BI__builtin_neon_vmul_lane_v:
8745   case NEON::BI__builtin_neon_vmul_laneq_v: {
8746     // v1f64 vmul_lane should be mapped to Neon scalar mul lane
8747     bool Quad = false;
8748     if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v)
8749       Quad = true;
8750     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
8751     llvm::Type *VTy = GetNeonType(this,
8752       NeonTypeFlags(NeonTypeFlags::Float64, false, Quad));
8753     Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
8754     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
8755     Value *Result = Builder.CreateFMul(Ops[0], Ops[1]);
8756     return Builder.CreateBitCast(Result, Ty);
8757   }
8758   case NEON::BI__builtin_neon_vnegd_s64:
8759     return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd");
8760   case NEON::BI__builtin_neon_vnegh_f16:
8761     return Builder.CreateFNeg(EmitScalarExpr(E->getArg(0)), "vnegh");
8762   case NEON::BI__builtin_neon_vpmaxnm_v:
8763   case NEON::BI__builtin_neon_vpmaxnmq_v: {
8764     Int = Intrinsic::aarch64_neon_fmaxnmp;
8765     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm");
8766   }
8767   case NEON::BI__builtin_neon_vpminnm_v:
8768   case NEON::BI__builtin_neon_vpminnmq_v: {
8769     Int = Intrinsic::aarch64_neon_fminnmp;
8770     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm");
8771   }
8772   case NEON::BI__builtin_neon_vsqrth_f16: {
8773     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8774     Int = Intrinsic::sqrt;
8775     return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vsqrt");
8776   }
8777   case NEON::BI__builtin_neon_vsqrt_v:
8778   case NEON::BI__builtin_neon_vsqrtq_v: {
8779     Int = Intrinsic::sqrt;
8780     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
8781     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt");
8782   }
8783   case NEON::BI__builtin_neon_vrbit_v:
8784   case NEON::BI__builtin_neon_vrbitq_v: {
8785     Int = Intrinsic::aarch64_neon_rbit;
8786     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit");
8787   }
8788   case NEON::BI__builtin_neon_vaddv_u8:
8789     // FIXME: These are handled by the AArch64 scalar code.
8790     usgn = true;
8791     LLVM_FALLTHROUGH;
8792   case NEON::BI__builtin_neon_vaddv_s8: {
8793     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
8794     Ty = Int32Ty;
8795     VTy = llvm::VectorType::get(Int8Ty, 8);
8796     llvm::Type *Tys[2] = { Ty, VTy };
8797     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8798     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
8799     return Builder.CreateTrunc(Ops[0], Int8Ty);
8800   }
8801   case NEON::BI__builtin_neon_vaddv_u16:
8802     usgn = true;
8803     LLVM_FALLTHROUGH;
8804   case NEON::BI__builtin_neon_vaddv_s16: {
8805     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
8806     Ty = Int32Ty;
8807     VTy = llvm::VectorType::get(Int16Ty, 4);
8808     llvm::Type *Tys[2] = { Ty, VTy };
8809     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8810     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
8811     return Builder.CreateTrunc(Ops[0], Int16Ty);
8812   }
8813   case NEON::BI__builtin_neon_vaddvq_u8:
8814     usgn = true;
8815     LLVM_FALLTHROUGH;
8816   case NEON::BI__builtin_neon_vaddvq_s8: {
8817     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
8818     Ty = Int32Ty;
8819     VTy = llvm::VectorType::get(Int8Ty, 16);
8820     llvm::Type *Tys[2] = { Ty, VTy };
8821     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8822     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
8823     return Builder.CreateTrunc(Ops[0], Int8Ty);
8824   }
8825   case NEON::BI__builtin_neon_vaddvq_u16:
8826     usgn = true;
8827     LLVM_FALLTHROUGH;
8828   case NEON::BI__builtin_neon_vaddvq_s16: {
8829     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
8830     Ty = Int32Ty;
8831     VTy = llvm::VectorType::get(Int16Ty, 8);
8832     llvm::Type *Tys[2] = { Ty, VTy };
8833     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8834     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
8835     return Builder.CreateTrunc(Ops[0], Int16Ty);
8836   }
8837   case NEON::BI__builtin_neon_vmaxv_u8: {
8838     Int = Intrinsic::aarch64_neon_umaxv;
8839     Ty = Int32Ty;
8840     VTy = llvm::VectorType::get(Int8Ty, 8);
8841     llvm::Type *Tys[2] = { Ty, VTy };
8842     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8843     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8844     return Builder.CreateTrunc(Ops[0], Int8Ty);
8845   }
8846   case NEON::BI__builtin_neon_vmaxv_u16: {
8847     Int = Intrinsic::aarch64_neon_umaxv;
8848     Ty = Int32Ty;
8849     VTy = llvm::VectorType::get(Int16Ty, 4);
8850     llvm::Type *Tys[2] = { Ty, VTy };
8851     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8852     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8853     return Builder.CreateTrunc(Ops[0], Int16Ty);
8854   }
8855   case NEON::BI__builtin_neon_vmaxvq_u8: {
8856     Int = Intrinsic::aarch64_neon_umaxv;
8857     Ty = Int32Ty;
8858     VTy = llvm::VectorType::get(Int8Ty, 16);
8859     llvm::Type *Tys[2] = { Ty, VTy };
8860     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8861     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8862     return Builder.CreateTrunc(Ops[0], Int8Ty);
8863   }
8864   case NEON::BI__builtin_neon_vmaxvq_u16: {
8865     Int = Intrinsic::aarch64_neon_umaxv;
8866     Ty = Int32Ty;
8867     VTy = llvm::VectorType::get(Int16Ty, 8);
8868     llvm::Type *Tys[2] = { Ty, VTy };
8869     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8870     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8871     return Builder.CreateTrunc(Ops[0], Int16Ty);
8872   }
8873   case NEON::BI__builtin_neon_vmaxv_s8: {
8874     Int = Intrinsic::aarch64_neon_smaxv;
8875     Ty = Int32Ty;
8876     VTy = llvm::VectorType::get(Int8Ty, 8);
8877     llvm::Type *Tys[2] = { Ty, VTy };
8878     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8879     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8880     return Builder.CreateTrunc(Ops[0], Int8Ty);
8881   }
8882   case NEON::BI__builtin_neon_vmaxv_s16: {
8883     Int = Intrinsic::aarch64_neon_smaxv;
8884     Ty = Int32Ty;
8885     VTy = llvm::VectorType::get(Int16Ty, 4);
8886     llvm::Type *Tys[2] = { Ty, VTy };
8887     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8888     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8889     return Builder.CreateTrunc(Ops[0], Int16Ty);
8890   }
8891   case NEON::BI__builtin_neon_vmaxvq_s8: {
8892     Int = Intrinsic::aarch64_neon_smaxv;
8893     Ty = Int32Ty;
8894     VTy = llvm::VectorType::get(Int8Ty, 16);
8895     llvm::Type *Tys[2] = { Ty, VTy };
8896     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8897     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8898     return Builder.CreateTrunc(Ops[0], Int8Ty);
8899   }
8900   case NEON::BI__builtin_neon_vmaxvq_s16: {
8901     Int = Intrinsic::aarch64_neon_smaxv;
8902     Ty = Int32Ty;
8903     VTy = llvm::VectorType::get(Int16Ty, 8);
8904     llvm::Type *Tys[2] = { Ty, VTy };
8905     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8906     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8907     return Builder.CreateTrunc(Ops[0], Int16Ty);
8908   }
8909   case NEON::BI__builtin_neon_vmaxv_f16: {
8910     Int = Intrinsic::aarch64_neon_fmaxv;
8911     Ty = HalfTy;
8912     VTy = llvm::VectorType::get(HalfTy, 4);
8913     llvm::Type *Tys[2] = { Ty, VTy };
8914     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8915     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8916     return Builder.CreateTrunc(Ops[0], HalfTy);
8917   }
8918   case NEON::BI__builtin_neon_vmaxvq_f16: {
8919     Int = Intrinsic::aarch64_neon_fmaxv;
8920     Ty = HalfTy;
8921     VTy = llvm::VectorType::get(HalfTy, 8);
8922     llvm::Type *Tys[2] = { Ty, VTy };
8923     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8924     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
8925     return Builder.CreateTrunc(Ops[0], HalfTy);
8926   }
8927   case NEON::BI__builtin_neon_vminv_u8: {
8928     Int = Intrinsic::aarch64_neon_uminv;
8929     Ty = Int32Ty;
8930     VTy = llvm::VectorType::get(Int8Ty, 8);
8931     llvm::Type *Tys[2] = { Ty, VTy };
8932     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8933     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8934     return Builder.CreateTrunc(Ops[0], Int8Ty);
8935   }
8936   case NEON::BI__builtin_neon_vminv_u16: {
8937     Int = Intrinsic::aarch64_neon_uminv;
8938     Ty = Int32Ty;
8939     VTy = llvm::VectorType::get(Int16Ty, 4);
8940     llvm::Type *Tys[2] = { Ty, VTy };
8941     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8942     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8943     return Builder.CreateTrunc(Ops[0], Int16Ty);
8944   }
8945   case NEON::BI__builtin_neon_vminvq_u8: {
8946     Int = Intrinsic::aarch64_neon_uminv;
8947     Ty = Int32Ty;
8948     VTy = llvm::VectorType::get(Int8Ty, 16);
8949     llvm::Type *Tys[2] = { Ty, VTy };
8950     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8951     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8952     return Builder.CreateTrunc(Ops[0], Int8Ty);
8953   }
8954   case NEON::BI__builtin_neon_vminvq_u16: {
8955     Int = Intrinsic::aarch64_neon_uminv;
8956     Ty = Int32Ty;
8957     VTy = llvm::VectorType::get(Int16Ty, 8);
8958     llvm::Type *Tys[2] = { Ty, VTy };
8959     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8960     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8961     return Builder.CreateTrunc(Ops[0], Int16Ty);
8962   }
8963   case NEON::BI__builtin_neon_vminv_s8: {
8964     Int = Intrinsic::aarch64_neon_sminv;
8965     Ty = Int32Ty;
8966     VTy = llvm::VectorType::get(Int8Ty, 8);
8967     llvm::Type *Tys[2] = { Ty, VTy };
8968     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8969     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8970     return Builder.CreateTrunc(Ops[0], Int8Ty);
8971   }
8972   case NEON::BI__builtin_neon_vminv_s16: {
8973     Int = Intrinsic::aarch64_neon_sminv;
8974     Ty = Int32Ty;
8975     VTy = llvm::VectorType::get(Int16Ty, 4);
8976     llvm::Type *Tys[2] = { Ty, VTy };
8977     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8978     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8979     return Builder.CreateTrunc(Ops[0], Int16Ty);
8980   }
8981   case NEON::BI__builtin_neon_vminvq_s8: {
8982     Int = Intrinsic::aarch64_neon_sminv;
8983     Ty = Int32Ty;
8984     VTy = llvm::VectorType::get(Int8Ty, 16);
8985     llvm::Type *Tys[2] = { Ty, VTy };
8986     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8987     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8988     return Builder.CreateTrunc(Ops[0], Int8Ty);
8989   }
8990   case NEON::BI__builtin_neon_vminvq_s16: {
8991     Int = Intrinsic::aarch64_neon_sminv;
8992     Ty = Int32Ty;
8993     VTy = llvm::VectorType::get(Int16Ty, 8);
8994     llvm::Type *Tys[2] = { Ty, VTy };
8995     Ops.push_back(EmitScalarExpr(E->getArg(0)));
8996     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
8997     return Builder.CreateTrunc(Ops[0], Int16Ty);
8998   }
8999   case NEON::BI__builtin_neon_vminv_f16: {
9000     Int = Intrinsic::aarch64_neon_fminv;
9001     Ty = HalfTy;
9002     VTy = llvm::VectorType::get(HalfTy, 4);
9003     llvm::Type *Tys[2] = { Ty, VTy };
9004     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9005     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
9006     return Builder.CreateTrunc(Ops[0], HalfTy);
9007   }
9008   case NEON::BI__builtin_neon_vminvq_f16: {
9009     Int = Intrinsic::aarch64_neon_fminv;
9010     Ty = HalfTy;
9011     VTy = llvm::VectorType::get(HalfTy, 8);
9012     llvm::Type *Tys[2] = { Ty, VTy };
9013     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9014     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
9015     return Builder.CreateTrunc(Ops[0], HalfTy);
9016   }
9017   case NEON::BI__builtin_neon_vmaxnmv_f16: {
9018     Int = Intrinsic::aarch64_neon_fmaxnmv;
9019     Ty = HalfTy;
9020     VTy = llvm::VectorType::get(HalfTy, 4);
9021     llvm::Type *Tys[2] = { Ty, VTy };
9022     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9023     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
9024     return Builder.CreateTrunc(Ops[0], HalfTy);
9025   }
9026   case NEON::BI__builtin_neon_vmaxnmvq_f16: {
9027     Int = Intrinsic::aarch64_neon_fmaxnmv;
9028     Ty = HalfTy;
9029     VTy = llvm::VectorType::get(HalfTy, 8);
9030     llvm::Type *Tys[2] = { Ty, VTy };
9031     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9032     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
9033     return Builder.CreateTrunc(Ops[0], HalfTy);
9034   }
9035   case NEON::BI__builtin_neon_vminnmv_f16: {
9036     Int = Intrinsic::aarch64_neon_fminnmv;
9037     Ty = HalfTy;
9038     VTy = llvm::VectorType::get(HalfTy, 4);
9039     llvm::Type *Tys[2] = { Ty, VTy };
9040     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9041     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
9042     return Builder.CreateTrunc(Ops[0], HalfTy);
9043   }
9044   case NEON::BI__builtin_neon_vminnmvq_f16: {
9045     Int = Intrinsic::aarch64_neon_fminnmv;
9046     Ty = HalfTy;
9047     VTy = llvm::VectorType::get(HalfTy, 8);
9048     llvm::Type *Tys[2] = { Ty, VTy };
9049     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9050     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
9051     return Builder.CreateTrunc(Ops[0], HalfTy);
9052   }
9053   case NEON::BI__builtin_neon_vmul_n_f64: {
9054     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
9055     Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy);
9056     return Builder.CreateFMul(Ops[0], RHS);
9057   }
9058   case NEON::BI__builtin_neon_vaddlv_u8: {
9059     Int = Intrinsic::aarch64_neon_uaddlv;
9060     Ty = Int32Ty;
9061     VTy = llvm::VectorType::get(Int8Ty, 8);
9062     llvm::Type *Tys[2] = { Ty, VTy };
9063     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9064     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9065     return Builder.CreateTrunc(Ops[0], Int16Ty);
9066   }
9067   case NEON::BI__builtin_neon_vaddlv_u16: {
9068     Int = Intrinsic::aarch64_neon_uaddlv;
9069     Ty = Int32Ty;
9070     VTy = llvm::VectorType::get(Int16Ty, 4);
9071     llvm::Type *Tys[2] = { Ty, VTy };
9072     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9073     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9074   }
9075   case NEON::BI__builtin_neon_vaddlvq_u8: {
9076     Int = Intrinsic::aarch64_neon_uaddlv;
9077     Ty = Int32Ty;
9078     VTy = llvm::VectorType::get(Int8Ty, 16);
9079     llvm::Type *Tys[2] = { Ty, VTy };
9080     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9081     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9082     return Builder.CreateTrunc(Ops[0], Int16Ty);
9083   }
9084   case NEON::BI__builtin_neon_vaddlvq_u16: {
9085     Int = Intrinsic::aarch64_neon_uaddlv;
9086     Ty = Int32Ty;
9087     VTy = llvm::VectorType::get(Int16Ty, 8);
9088     llvm::Type *Tys[2] = { Ty, VTy };
9089     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9090     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9091   }
9092   case NEON::BI__builtin_neon_vaddlv_s8: {
9093     Int = Intrinsic::aarch64_neon_saddlv;
9094     Ty = Int32Ty;
9095     VTy = llvm::VectorType::get(Int8Ty, 8);
9096     llvm::Type *Tys[2] = { Ty, VTy };
9097     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9098     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9099     return Builder.CreateTrunc(Ops[0], Int16Ty);
9100   }
9101   case NEON::BI__builtin_neon_vaddlv_s16: {
9102     Int = Intrinsic::aarch64_neon_saddlv;
9103     Ty = Int32Ty;
9104     VTy = llvm::VectorType::get(Int16Ty, 4);
9105     llvm::Type *Tys[2] = { Ty, VTy };
9106     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9107     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9108   }
9109   case NEON::BI__builtin_neon_vaddlvq_s8: {
9110     Int = Intrinsic::aarch64_neon_saddlv;
9111     Ty = Int32Ty;
9112     VTy = llvm::VectorType::get(Int8Ty, 16);
9113     llvm::Type *Tys[2] = { Ty, VTy };
9114     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9115     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9116     return Builder.CreateTrunc(Ops[0], Int16Ty);
9117   }
9118   case NEON::BI__builtin_neon_vaddlvq_s16: {
9119     Int = Intrinsic::aarch64_neon_saddlv;
9120     Ty = Int32Ty;
9121     VTy = llvm::VectorType::get(Int16Ty, 8);
9122     llvm::Type *Tys[2] = { Ty, VTy };
9123     Ops.push_back(EmitScalarExpr(E->getArg(0)));
9124     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
9125   }
9126   case NEON::BI__builtin_neon_vsri_n_v:
9127   case NEON::BI__builtin_neon_vsriq_n_v: {
9128     Int = Intrinsic::aarch64_neon_vsri;
9129     llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
9130     return EmitNeonCall(Intrin, Ops, "vsri_n");
9131   }
9132   case NEON::BI__builtin_neon_vsli_n_v:
9133   case NEON::BI__builtin_neon_vsliq_n_v: {
9134     Int = Intrinsic::aarch64_neon_vsli;
9135     llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
9136     return EmitNeonCall(Intrin, Ops, "vsli_n");
9137   }
9138   case NEON::BI__builtin_neon_vsra_n_v:
9139   case NEON::BI__builtin_neon_vsraq_n_v:
9140     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9141     Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
9142     return Builder.CreateAdd(Ops[0], Ops[1]);
9143   case NEON::BI__builtin_neon_vrsra_n_v:
9144   case NEON::BI__builtin_neon_vrsraq_n_v: {
9145     Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl;
9146     SmallVector<llvm::Value*,2> TmpOps;
9147     TmpOps.push_back(Ops[1]);
9148     TmpOps.push_back(Ops[2]);
9149     Function* F = CGM.getIntrinsic(Int, Ty);
9150     llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true);
9151     Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
9152     return Builder.CreateAdd(Ops[0], tmp);
9153   }
9154   case NEON::BI__builtin_neon_vld1_v:
9155   case NEON::BI__builtin_neon_vld1q_v: {
9156     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
9157     auto Alignment = CharUnits::fromQuantity(
9158         BuiltinID == NEON::BI__builtin_neon_vld1_v ? 8 : 16);
9159     return Builder.CreateAlignedLoad(VTy, Ops[0], Alignment);
9160   }
9161   case NEON::BI__builtin_neon_vst1_v:
9162   case NEON::BI__builtin_neon_vst1q_v:
9163     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
9164     Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
9165     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9166   case NEON::BI__builtin_neon_vld1_lane_v:
9167   case NEON::BI__builtin_neon_vld1q_lane_v: {
9168     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9169     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
9170     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9171     auto Alignment = CharUnits::fromQuantity(
9172         BuiltinID == NEON::BI__builtin_neon_vld1_lane_v ? 8 : 16);
9173     Ops[0] =
9174         Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], Alignment);
9175     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane");
9176   }
9177   case NEON::BI__builtin_neon_vld1_dup_v:
9178   case NEON::BI__builtin_neon_vld1q_dup_v: {
9179     Value *V = UndefValue::get(Ty);
9180     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
9181     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9182     auto Alignment = CharUnits::fromQuantity(
9183         BuiltinID == NEON::BI__builtin_neon_vld1_dup_v ? 8 : 16);
9184     Ops[0] =
9185         Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], Alignment);
9186     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
9187     Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
9188     return EmitNeonSplat(Ops[0], CI);
9189   }
9190   case NEON::BI__builtin_neon_vst1_lane_v:
9191   case NEON::BI__builtin_neon_vst1q_lane_v:
9192     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9193     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
9194     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
9195     return Builder.CreateDefaultAlignedStore(Ops[1],
9196                                              Builder.CreateBitCast(Ops[0], Ty));
9197   case NEON::BI__builtin_neon_vld2_v:
9198   case NEON::BI__builtin_neon_vld2q_v: {
9199     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
9200     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
9201     llvm::Type *Tys[2] = { VTy, PTy };
9202     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys);
9203     Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
9204     Ops[0] = Builder.CreateBitCast(Ops[0],
9205                 llvm::PointerType::getUnqual(Ops[1]->getType()));
9206     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9207   }
9208   case NEON::BI__builtin_neon_vld3_v:
9209   case NEON::BI__builtin_neon_vld3q_v: {
9210     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
9211     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
9212     llvm::Type *Tys[2] = { VTy, PTy };
9213     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys);
9214     Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
9215     Ops[0] = Builder.CreateBitCast(Ops[0],
9216                 llvm::PointerType::getUnqual(Ops[1]->getType()));
9217     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9218   }
9219   case NEON::BI__builtin_neon_vld4_v:
9220   case NEON::BI__builtin_neon_vld4q_v: {
9221     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
9222     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
9223     llvm::Type *Tys[2] = { VTy, PTy };
9224     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys);
9225     Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
9226     Ops[0] = Builder.CreateBitCast(Ops[0],
9227                 llvm::PointerType::getUnqual(Ops[1]->getType()));
9228     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9229   }
9230   case NEON::BI__builtin_neon_vld2_dup_v:
9231   case NEON::BI__builtin_neon_vld2q_dup_v: {
9232     llvm::Type *PTy =
9233       llvm::PointerType::getUnqual(VTy->getElementType());
9234     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
9235     llvm::Type *Tys[2] = { VTy, PTy };
9236     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys);
9237     Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
9238     Ops[0] = Builder.CreateBitCast(Ops[0],
9239                 llvm::PointerType::getUnqual(Ops[1]->getType()));
9240     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9241   }
9242   case NEON::BI__builtin_neon_vld3_dup_v:
9243   case NEON::BI__builtin_neon_vld3q_dup_v: {
9244     llvm::Type *PTy =
9245       llvm::PointerType::getUnqual(VTy->getElementType());
9246     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
9247     llvm::Type *Tys[2] = { VTy, PTy };
9248     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys);
9249     Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
9250     Ops[0] = Builder.CreateBitCast(Ops[0],
9251                 llvm::PointerType::getUnqual(Ops[1]->getType()));
9252     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9253   }
9254   case NEON::BI__builtin_neon_vld4_dup_v:
9255   case NEON::BI__builtin_neon_vld4q_dup_v: {
9256     llvm::Type *PTy =
9257       llvm::PointerType::getUnqual(VTy->getElementType());
9258     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
9259     llvm::Type *Tys[2] = { VTy, PTy };
9260     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys);
9261     Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
9262     Ops[0] = Builder.CreateBitCast(Ops[0],
9263                 llvm::PointerType::getUnqual(Ops[1]->getType()));
9264     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9265   }
9266   case NEON::BI__builtin_neon_vld2_lane_v:
9267   case NEON::BI__builtin_neon_vld2q_lane_v: {
9268     llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
9269     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys);
9270     Ops.push_back(Ops[1]);
9271     Ops.erase(Ops.begin()+1);
9272     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9273     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
9274     Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
9275     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane");
9276     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
9277     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9278     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9279   }
9280   case NEON::BI__builtin_neon_vld3_lane_v:
9281   case NEON::BI__builtin_neon_vld3q_lane_v: {
9282     llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
9283     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys);
9284     Ops.push_back(Ops[1]);
9285     Ops.erase(Ops.begin()+1);
9286     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9287     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
9288     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
9289     Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty);
9290     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
9291     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
9292     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9293     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9294   }
9295   case NEON::BI__builtin_neon_vld4_lane_v:
9296   case NEON::BI__builtin_neon_vld4q_lane_v: {
9297     llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
9298     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys);
9299     Ops.push_back(Ops[1]);
9300     Ops.erase(Ops.begin()+1);
9301     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9302     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
9303     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
9304     Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
9305     Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty);
9306     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane");
9307     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
9308     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9309     return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
9310   }
9311   case NEON::BI__builtin_neon_vst2_v:
9312   case NEON::BI__builtin_neon_vst2q_v: {
9313     Ops.push_back(Ops[0]);
9314     Ops.erase(Ops.begin());
9315     llvm::Type *Tys[2] = { VTy, Ops[2]->getType() };
9316     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys),
9317                         Ops, "");
9318   }
9319   case NEON::BI__builtin_neon_vst2_lane_v:
9320   case NEON::BI__builtin_neon_vst2q_lane_v: {
9321     Ops.push_back(Ops[0]);
9322     Ops.erase(Ops.begin());
9323     Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
9324     llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
9325     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys),
9326                         Ops, "");
9327   }
9328   case NEON::BI__builtin_neon_vst3_v:
9329   case NEON::BI__builtin_neon_vst3q_v: {
9330     Ops.push_back(Ops[0]);
9331     Ops.erase(Ops.begin());
9332     llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
9333     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys),
9334                         Ops, "");
9335   }
9336   case NEON::BI__builtin_neon_vst3_lane_v:
9337   case NEON::BI__builtin_neon_vst3q_lane_v: {
9338     Ops.push_back(Ops[0]);
9339     Ops.erase(Ops.begin());
9340     Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
9341     llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
9342     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys),
9343                         Ops, "");
9344   }
9345   case NEON::BI__builtin_neon_vst4_v:
9346   case NEON::BI__builtin_neon_vst4q_v: {
9347     Ops.push_back(Ops[0]);
9348     Ops.erase(Ops.begin());
9349     llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
9350     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys),
9351                         Ops, "");
9352   }
9353   case NEON::BI__builtin_neon_vst4_lane_v:
9354   case NEON::BI__builtin_neon_vst4q_lane_v: {
9355     Ops.push_back(Ops[0]);
9356     Ops.erase(Ops.begin());
9357     Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty);
9358     llvm::Type *Tys[2] = { VTy, Ops[5]->getType() };
9359     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys),
9360                         Ops, "");
9361   }
9362   case NEON::BI__builtin_neon_vtrn_v:
9363   case NEON::BI__builtin_neon_vtrnq_v: {
9364     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
9365     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9366     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
9367     Value *SV = nullptr;
9368 
9369     for (unsigned vi = 0; vi != 2; ++vi) {
9370       SmallVector<uint32_t, 16> Indices;
9371       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
9372         Indices.push_back(i+vi);
9373         Indices.push_back(i+e+vi);
9374       }
9375       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
9376       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn");
9377       SV = Builder.CreateDefaultAlignedStore(SV, Addr);
9378     }
9379     return SV;
9380   }
9381   case NEON::BI__builtin_neon_vuzp_v:
9382   case NEON::BI__builtin_neon_vuzpq_v: {
9383     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
9384     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9385     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
9386     Value *SV = nullptr;
9387 
9388     for (unsigned vi = 0; vi != 2; ++vi) {
9389       SmallVector<uint32_t, 16> Indices;
9390       for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
9391         Indices.push_back(2*i+vi);
9392 
9393       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
9394       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp");
9395       SV = Builder.CreateDefaultAlignedStore(SV, Addr);
9396     }
9397     return SV;
9398   }
9399   case NEON::BI__builtin_neon_vzip_v:
9400   case NEON::BI__builtin_neon_vzipq_v: {
9401     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
9402     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9403     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
9404     Value *SV = nullptr;
9405 
9406     for (unsigned vi = 0; vi != 2; ++vi) {
9407       SmallVector<uint32_t, 16> Indices;
9408       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
9409         Indices.push_back((i + vi*e) >> 1);
9410         Indices.push_back(((i + vi*e) >> 1)+e);
9411       }
9412       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
9413       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip");
9414       SV = Builder.CreateDefaultAlignedStore(SV, Addr);
9415     }
9416     return SV;
9417   }
9418   case NEON::BI__builtin_neon_vqtbl1q_v: {
9419     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty),
9420                         Ops, "vtbl1");
9421   }
9422   case NEON::BI__builtin_neon_vqtbl2q_v: {
9423     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty),
9424                         Ops, "vtbl2");
9425   }
9426   case NEON::BI__builtin_neon_vqtbl3q_v: {
9427     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty),
9428                         Ops, "vtbl3");
9429   }
9430   case NEON::BI__builtin_neon_vqtbl4q_v: {
9431     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty),
9432                         Ops, "vtbl4");
9433   }
9434   case NEON::BI__builtin_neon_vqtbx1q_v: {
9435     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty),
9436                         Ops, "vtbx1");
9437   }
9438   case NEON::BI__builtin_neon_vqtbx2q_v: {
9439     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty),
9440                         Ops, "vtbx2");
9441   }
9442   case NEON::BI__builtin_neon_vqtbx3q_v: {
9443     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty),
9444                         Ops, "vtbx3");
9445   }
9446   case NEON::BI__builtin_neon_vqtbx4q_v: {
9447     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty),
9448                         Ops, "vtbx4");
9449   }
9450   case NEON::BI__builtin_neon_vsqadd_v:
9451   case NEON::BI__builtin_neon_vsqaddq_v: {
9452     Int = Intrinsic::aarch64_neon_usqadd;
9453     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd");
9454   }
9455   case NEON::BI__builtin_neon_vuqadd_v:
9456   case NEON::BI__builtin_neon_vuqaddq_v: {
9457     Int = Intrinsic::aarch64_neon_suqadd;
9458     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd");
9459   }
9460   }
9461 }
9462 
9463 Value *CodeGenFunction::EmitBPFBuiltinExpr(unsigned BuiltinID,
9464                                            const CallExpr *E) {
9465   assert(BuiltinID == BPF::BI__builtin_preserve_field_info &&
9466          "unexpected ARM builtin");
9467 
9468   const Expr *Arg = E->getArg(0);
9469   bool IsBitField = Arg->IgnoreParens()->getObjectKind() == OK_BitField;
9470 
9471   if (!getDebugInfo()) {
9472     CGM.Error(E->getExprLoc(), "using builtin_preserve_field_info() without -g");
9473     return IsBitField ? EmitLValue(Arg).getBitFieldPointer()
9474                       : EmitLValue(Arg).getPointer(*this);
9475   }
9476 
9477   // Enable underlying preserve_*_access_index() generation.
9478   bool OldIsInPreservedAIRegion = IsInPreservedAIRegion;
9479   IsInPreservedAIRegion = true;
9480   Value *FieldAddr = IsBitField ? EmitLValue(Arg).getBitFieldPointer()
9481                                 : EmitLValue(Arg).getPointer(*this);
9482   IsInPreservedAIRegion = OldIsInPreservedAIRegion;
9483 
9484   ConstantInt *C = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
9485   Value *InfoKind = ConstantInt::get(Int64Ty, C->getSExtValue());
9486 
9487   // Built the IR for the preserve_field_info intrinsic.
9488   llvm::Function *FnGetFieldInfo = llvm::Intrinsic::getDeclaration(
9489       &CGM.getModule(), llvm::Intrinsic::bpf_preserve_field_info,
9490       {FieldAddr->getType()});
9491   return Builder.CreateCall(FnGetFieldInfo, {FieldAddr, InfoKind});
9492 }
9493 
9494 llvm::Value *CodeGenFunction::
9495 BuildVector(ArrayRef<llvm::Value*> Ops) {
9496   assert((Ops.size() & (Ops.size() - 1)) == 0 &&
9497          "Not a power-of-two sized vector!");
9498   bool AllConstants = true;
9499   for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
9500     AllConstants &= isa<Constant>(Ops[i]);
9501 
9502   // If this is a constant vector, create a ConstantVector.
9503   if (AllConstants) {
9504     SmallVector<llvm::Constant*, 16> CstOps;
9505     for (unsigned i = 0, e = Ops.size(); i != e; ++i)
9506       CstOps.push_back(cast<Constant>(Ops[i]));
9507     return llvm::ConstantVector::get(CstOps);
9508   }
9509 
9510   // Otherwise, insertelement the values to build the vector.
9511   Value *Result =
9512     llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size()));
9513 
9514   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
9515     Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i));
9516 
9517   return Result;
9518 }
9519 
9520 // Convert the mask from an integer type to a vector of i1.
9521 static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask,
9522                               unsigned NumElts) {
9523 
9524   llvm::VectorType *MaskTy = llvm::VectorType::get(CGF.Builder.getInt1Ty(),
9525                          cast<IntegerType>(Mask->getType())->getBitWidth());
9526   Value *MaskVec = CGF.Builder.CreateBitCast(Mask, MaskTy);
9527 
9528   // If we have less than 8 elements, then the starting mask was an i8 and
9529   // we need to extract down to the right number of elements.
9530   if (NumElts < 8) {
9531     uint32_t Indices[4];
9532     for (unsigned i = 0; i != NumElts; ++i)
9533       Indices[i] = i;
9534     MaskVec = CGF.Builder.CreateShuffleVector(MaskVec, MaskVec,
9535                                              makeArrayRef(Indices, NumElts),
9536                                              "extract");
9537   }
9538   return MaskVec;
9539 }
9540 
9541 static Value *EmitX86MaskedStore(CodeGenFunction &CGF,
9542                                  ArrayRef<Value *> Ops,
9543                                  unsigned Align) {
9544   // Cast the pointer to right type.
9545   Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
9546                                llvm::PointerType::getUnqual(Ops[1]->getType()));
9547 
9548   Value *MaskVec = getMaskVecValue(CGF, Ops[2],
9549                                    Ops[1]->getType()->getVectorNumElements());
9550 
9551   return CGF.Builder.CreateMaskedStore(Ops[1], Ptr, Align, MaskVec);
9552 }
9553 
9554 static Value *EmitX86MaskedLoad(CodeGenFunction &CGF,
9555                                 ArrayRef<Value *> Ops, unsigned Align) {
9556   // Cast the pointer to right type.
9557   Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
9558                                llvm::PointerType::getUnqual(Ops[1]->getType()));
9559 
9560   Value *MaskVec = getMaskVecValue(CGF, Ops[2],
9561                                    Ops[1]->getType()->getVectorNumElements());
9562 
9563   return CGF.Builder.CreateMaskedLoad(Ptr, Align, MaskVec, Ops[1]);
9564 }
9565 
9566 static Value *EmitX86ExpandLoad(CodeGenFunction &CGF,
9567                                 ArrayRef<Value *> Ops) {
9568   llvm::Type *ResultTy = Ops[1]->getType();
9569   llvm::Type *PtrTy = ResultTy->getVectorElementType();
9570 
9571   // Cast the pointer to element type.
9572   Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
9573                                          llvm::PointerType::getUnqual(PtrTy));
9574 
9575   Value *MaskVec = getMaskVecValue(CGF, Ops[2],
9576                                    ResultTy->getVectorNumElements());
9577 
9578   llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_expandload,
9579                                            ResultTy);
9580   return CGF.Builder.CreateCall(F, { Ptr, MaskVec, Ops[1] });
9581 }
9582 
9583 static Value *EmitX86CompressExpand(CodeGenFunction &CGF,
9584                                     ArrayRef<Value *> Ops,
9585                                     bool IsCompress) {
9586   llvm::Type *ResultTy = Ops[1]->getType();
9587 
9588   Value *MaskVec = getMaskVecValue(CGF, Ops[2],
9589                                    ResultTy->getVectorNumElements());
9590 
9591   Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress
9592                                  : Intrinsic::x86_avx512_mask_expand;
9593   llvm::Function *F = CGF.CGM.getIntrinsic(IID, ResultTy);
9594   return CGF.Builder.CreateCall(F, { Ops[0], Ops[1], MaskVec });
9595 }
9596 
9597 static Value *EmitX86CompressStore(CodeGenFunction &CGF,
9598                                    ArrayRef<Value *> Ops) {
9599   llvm::Type *ResultTy = Ops[1]->getType();
9600   llvm::Type *PtrTy = ResultTy->getVectorElementType();
9601 
9602   // Cast the pointer to element type.
9603   Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
9604                                          llvm::PointerType::getUnqual(PtrTy));
9605 
9606   Value *MaskVec = getMaskVecValue(CGF, Ops[2],
9607                                    ResultTy->getVectorNumElements());
9608 
9609   llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_compressstore,
9610                                            ResultTy);
9611   return CGF.Builder.CreateCall(F, { Ops[1], Ptr, MaskVec });
9612 }
9613 
9614 static Value *EmitX86MaskLogic(CodeGenFunction &CGF, Instruction::BinaryOps Opc,
9615                               ArrayRef<Value *> Ops,
9616                               bool InvertLHS = false) {
9617   unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
9618   Value *LHS = getMaskVecValue(CGF, Ops[0], NumElts);
9619   Value *RHS = getMaskVecValue(CGF, Ops[1], NumElts);
9620 
9621   if (InvertLHS)
9622     LHS = CGF.Builder.CreateNot(LHS);
9623 
9624   return CGF.Builder.CreateBitCast(CGF.Builder.CreateBinOp(Opc, LHS, RHS),
9625                                    Ops[0]->getType());
9626 }
9627 
9628 static Value *EmitX86FunnelShift(CodeGenFunction &CGF, Value *Op0, Value *Op1,
9629                                  Value *Amt, bool IsRight) {
9630   llvm::Type *Ty = Op0->getType();
9631 
9632   // Amount may be scalar immediate, in which case create a splat vector.
9633   // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
9634   // we only care about the lowest log2 bits anyway.
9635   if (Amt->getType() != Ty) {
9636     unsigned NumElts = Ty->getVectorNumElements();
9637     Amt = CGF.Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
9638     Amt = CGF.Builder.CreateVectorSplat(NumElts, Amt);
9639   }
9640 
9641   unsigned IID = IsRight ? Intrinsic::fshr : Intrinsic::fshl;
9642   Function *F = CGF.CGM.getIntrinsic(IID, Ty);
9643   return CGF.Builder.CreateCall(F, {Op0, Op1, Amt});
9644 }
9645 
9646 static Value *EmitX86vpcom(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
9647                            bool IsSigned) {
9648   Value *Op0 = Ops[0];
9649   Value *Op1 = Ops[1];
9650   llvm::Type *Ty = Op0->getType();
9651   uint64_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
9652 
9653   CmpInst::Predicate Pred;
9654   switch (Imm) {
9655   case 0x0:
9656     Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
9657     break;
9658   case 0x1:
9659     Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
9660     break;
9661   case 0x2:
9662     Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
9663     break;
9664   case 0x3:
9665     Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
9666     break;
9667   case 0x4:
9668     Pred = ICmpInst::ICMP_EQ;
9669     break;
9670   case 0x5:
9671     Pred = ICmpInst::ICMP_NE;
9672     break;
9673   case 0x6:
9674     return llvm::Constant::getNullValue(Ty); // FALSE
9675   case 0x7:
9676     return llvm::Constant::getAllOnesValue(Ty); // TRUE
9677   default:
9678     llvm_unreachable("Unexpected XOP vpcom/vpcomu predicate");
9679   }
9680 
9681   Value *Cmp = CGF.Builder.CreateICmp(Pred, Op0, Op1);
9682   Value *Res = CGF.Builder.CreateSExt(Cmp, Ty);
9683   return Res;
9684 }
9685 
9686 static Value *EmitX86Select(CodeGenFunction &CGF,
9687                             Value *Mask, Value *Op0, Value *Op1) {
9688 
9689   // If the mask is all ones just return first argument.
9690   if (const auto *C = dyn_cast<Constant>(Mask))
9691     if (C->isAllOnesValue())
9692       return Op0;
9693 
9694   Mask = getMaskVecValue(CGF, Mask, Op0->getType()->getVectorNumElements());
9695 
9696   return CGF.Builder.CreateSelect(Mask, Op0, Op1);
9697 }
9698 
9699 static Value *EmitX86ScalarSelect(CodeGenFunction &CGF,
9700                                   Value *Mask, Value *Op0, Value *Op1) {
9701   // If the mask is all ones just return first argument.
9702   if (const auto *C = dyn_cast<Constant>(Mask))
9703     if (C->isAllOnesValue())
9704       return Op0;
9705 
9706   llvm::VectorType *MaskTy =
9707     llvm::VectorType::get(CGF.Builder.getInt1Ty(),
9708                           Mask->getType()->getIntegerBitWidth());
9709   Mask = CGF.Builder.CreateBitCast(Mask, MaskTy);
9710   Mask = CGF.Builder.CreateExtractElement(Mask, (uint64_t)0);
9711   return CGF.Builder.CreateSelect(Mask, Op0, Op1);
9712 }
9713 
9714 static Value *EmitX86MaskedCompareResult(CodeGenFunction &CGF, Value *Cmp,
9715                                          unsigned NumElts, Value *MaskIn) {
9716   if (MaskIn) {
9717     const auto *C = dyn_cast<Constant>(MaskIn);
9718     if (!C || !C->isAllOnesValue())
9719       Cmp = CGF.Builder.CreateAnd(Cmp, getMaskVecValue(CGF, MaskIn, NumElts));
9720   }
9721 
9722   if (NumElts < 8) {
9723     uint32_t Indices[8];
9724     for (unsigned i = 0; i != NumElts; ++i)
9725       Indices[i] = i;
9726     for (unsigned i = NumElts; i != 8; ++i)
9727       Indices[i] = i % NumElts + NumElts;
9728     Cmp = CGF.Builder.CreateShuffleVector(
9729         Cmp, llvm::Constant::getNullValue(Cmp->getType()), Indices);
9730   }
9731 
9732   return CGF.Builder.CreateBitCast(Cmp,
9733                                    IntegerType::get(CGF.getLLVMContext(),
9734                                                     std::max(NumElts, 8U)));
9735 }
9736 
9737 static Value *EmitX86MaskedCompare(CodeGenFunction &CGF, unsigned CC,
9738                                    bool Signed, ArrayRef<Value *> Ops) {
9739   assert((Ops.size() == 2 || Ops.size() == 4) &&
9740          "Unexpected number of arguments");
9741   unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
9742   Value *Cmp;
9743 
9744   if (CC == 3) {
9745     Cmp = Constant::getNullValue(
9746                        llvm::VectorType::get(CGF.Builder.getInt1Ty(), NumElts));
9747   } else if (CC == 7) {
9748     Cmp = Constant::getAllOnesValue(
9749                        llvm::VectorType::get(CGF.Builder.getInt1Ty(), NumElts));
9750   } else {
9751     ICmpInst::Predicate Pred;
9752     switch (CC) {
9753     default: llvm_unreachable("Unknown condition code");
9754     case 0: Pred = ICmpInst::ICMP_EQ;  break;
9755     case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
9756     case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
9757     case 4: Pred = ICmpInst::ICMP_NE;  break;
9758     case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
9759     case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
9760     }
9761     Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]);
9762   }
9763 
9764   Value *MaskIn = nullptr;
9765   if (Ops.size() == 4)
9766     MaskIn = Ops[3];
9767 
9768   return EmitX86MaskedCompareResult(CGF, Cmp, NumElts, MaskIn);
9769 }
9770 
9771 static Value *EmitX86ConvertToMask(CodeGenFunction &CGF, Value *In) {
9772   Value *Zero = Constant::getNullValue(In->getType());
9773   return EmitX86MaskedCompare(CGF, 1, true, { In, Zero });
9774 }
9775 
9776 static Value *EmitX86ConvertIntToFp(CodeGenFunction &CGF,
9777                                     ArrayRef<Value *> Ops, bool IsSigned) {
9778   unsigned Rnd = cast<llvm::ConstantInt>(Ops[3])->getZExtValue();
9779   llvm::Type *Ty = Ops[1]->getType();
9780 
9781   Value *Res;
9782   if (Rnd != 4) {
9783     Intrinsic::ID IID = IsSigned ? Intrinsic::x86_avx512_sitofp_round
9784                                  : Intrinsic::x86_avx512_uitofp_round;
9785     Function *F = CGF.CGM.getIntrinsic(IID, { Ty, Ops[0]->getType() });
9786     Res = CGF.Builder.CreateCall(F, { Ops[0], Ops[3] });
9787   } else {
9788     Res = IsSigned ? CGF.Builder.CreateSIToFP(Ops[0], Ty)
9789                    : CGF.Builder.CreateUIToFP(Ops[0], Ty);
9790   }
9791 
9792   return EmitX86Select(CGF, Ops[2], Res, Ops[1]);
9793 }
9794 
9795 static Value *EmitX86Abs(CodeGenFunction &CGF, ArrayRef<Value *> Ops) {
9796 
9797   llvm::Type *Ty = Ops[0]->getType();
9798   Value *Zero = llvm::Constant::getNullValue(Ty);
9799   Value *Sub = CGF.Builder.CreateSub(Zero, Ops[0]);
9800   Value *Cmp = CGF.Builder.CreateICmp(ICmpInst::ICMP_SGT, Ops[0], Zero);
9801   Value *Res = CGF.Builder.CreateSelect(Cmp, Ops[0], Sub);
9802   return Res;
9803 }
9804 
9805 static Value *EmitX86MinMax(CodeGenFunction &CGF, ICmpInst::Predicate Pred,
9806                             ArrayRef<Value *> Ops) {
9807   Value *Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]);
9808   Value *Res = CGF.Builder.CreateSelect(Cmp, Ops[0], Ops[1]);
9809 
9810   assert(Ops.size() == 2);
9811   return Res;
9812 }
9813 
9814 // Lowers X86 FMA intrinsics to IR.
9815 static Value *EmitX86FMAExpr(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
9816                              unsigned BuiltinID, bool IsAddSub) {
9817 
9818   bool Subtract = false;
9819   Intrinsic::ID IID = Intrinsic::not_intrinsic;
9820   switch (BuiltinID) {
9821   default: break;
9822   case clang::X86::BI__builtin_ia32_vfmsubps512_mask3:
9823     Subtract = true;
9824     LLVM_FALLTHROUGH;
9825   case clang::X86::BI__builtin_ia32_vfmaddps512_mask:
9826   case clang::X86::BI__builtin_ia32_vfmaddps512_maskz:
9827   case clang::X86::BI__builtin_ia32_vfmaddps512_mask3:
9828     IID = llvm::Intrinsic::x86_avx512_vfmadd_ps_512; break;
9829   case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3:
9830     Subtract = true;
9831     LLVM_FALLTHROUGH;
9832   case clang::X86::BI__builtin_ia32_vfmaddpd512_mask:
9833   case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz:
9834   case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3:
9835     IID = llvm::Intrinsic::x86_avx512_vfmadd_pd_512; break;
9836   case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3:
9837     Subtract = true;
9838     LLVM_FALLTHROUGH;
9839   case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask:
9840   case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz:
9841   case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3:
9842     IID = llvm::Intrinsic::x86_avx512_vfmaddsub_ps_512;
9843     break;
9844   case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
9845     Subtract = true;
9846     LLVM_FALLTHROUGH;
9847   case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask:
9848   case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
9849   case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
9850     IID = llvm::Intrinsic::x86_avx512_vfmaddsub_pd_512;
9851     break;
9852   }
9853 
9854   Value *A = Ops[0];
9855   Value *B = Ops[1];
9856   Value *C = Ops[2];
9857 
9858   if (Subtract)
9859     C = CGF.Builder.CreateFNeg(C);
9860 
9861   Value *Res;
9862 
9863   // Only handle in case of _MM_FROUND_CUR_DIRECTION/4 (no rounding).
9864   if (IID != Intrinsic::not_intrinsic &&
9865       cast<llvm::ConstantInt>(Ops.back())->getZExtValue() != (uint64_t)4) {
9866     Function *Intr = CGF.CGM.getIntrinsic(IID);
9867     Res = CGF.Builder.CreateCall(Intr, {A, B, C, Ops.back() });
9868   } else {
9869     llvm::Type *Ty = A->getType();
9870     Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty);
9871     Res = CGF.Builder.CreateCall(FMA, {A, B, C} );
9872 
9873     if (IsAddSub) {
9874       // Negate even elts in C using a mask.
9875       unsigned NumElts = Ty->getVectorNumElements();
9876       SmallVector<uint32_t, 16> Indices(NumElts);
9877       for (unsigned i = 0; i != NumElts; ++i)
9878         Indices[i] = i + (i % 2) * NumElts;
9879 
9880       Value *NegC = CGF.Builder.CreateFNeg(C);
9881       Value *FMSub = CGF.Builder.CreateCall(FMA, {A, B, NegC} );
9882       Res = CGF.Builder.CreateShuffleVector(FMSub, Res, Indices);
9883     }
9884   }
9885 
9886   // Handle any required masking.
9887   Value *MaskFalseVal = nullptr;
9888   switch (BuiltinID) {
9889   case clang::X86::BI__builtin_ia32_vfmaddps512_mask:
9890   case clang::X86::BI__builtin_ia32_vfmaddpd512_mask:
9891   case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask:
9892   case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask:
9893     MaskFalseVal = Ops[0];
9894     break;
9895   case clang::X86::BI__builtin_ia32_vfmaddps512_maskz:
9896   case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz:
9897   case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz:
9898   case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
9899     MaskFalseVal = Constant::getNullValue(Ops[0]->getType());
9900     break;
9901   case clang::X86::BI__builtin_ia32_vfmsubps512_mask3:
9902   case clang::X86::BI__builtin_ia32_vfmaddps512_mask3:
9903   case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3:
9904   case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3:
9905   case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3:
9906   case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3:
9907   case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
9908   case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
9909     MaskFalseVal = Ops[2];
9910     break;
9911   }
9912 
9913   if (MaskFalseVal)
9914     return EmitX86Select(CGF, Ops[3], Res, MaskFalseVal);
9915 
9916   return Res;
9917 }
9918 
9919 static Value *
9920 EmitScalarFMAExpr(CodeGenFunction &CGF, MutableArrayRef<Value *> Ops,
9921                   Value *Upper, bool ZeroMask = false, unsigned PTIdx = 0,
9922                   bool NegAcc = false) {
9923   unsigned Rnd = 4;
9924   if (Ops.size() > 4)
9925     Rnd = cast<llvm::ConstantInt>(Ops[4])->getZExtValue();
9926 
9927   if (NegAcc)
9928     Ops[2] = CGF.Builder.CreateFNeg(Ops[2]);
9929 
9930   Ops[0] = CGF.Builder.CreateExtractElement(Ops[0], (uint64_t)0);
9931   Ops[1] = CGF.Builder.CreateExtractElement(Ops[1], (uint64_t)0);
9932   Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], (uint64_t)0);
9933   Value *Res;
9934   if (Rnd != 4) {
9935     Intrinsic::ID IID = Ops[0]->getType()->getPrimitiveSizeInBits() == 32 ?
9936                         Intrinsic::x86_avx512_vfmadd_f32 :
9937                         Intrinsic::x86_avx512_vfmadd_f64;
9938     Res = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID),
9939                                  {Ops[0], Ops[1], Ops[2], Ops[4]});
9940   } else {
9941     Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ops[0]->getType());
9942     Res = CGF.Builder.CreateCall(FMA, Ops.slice(0, 3));
9943   }
9944   // If we have more than 3 arguments, we need to do masking.
9945   if (Ops.size() > 3) {
9946     Value *PassThru = ZeroMask ? Constant::getNullValue(Res->getType())
9947                                : Ops[PTIdx];
9948 
9949     // If we negated the accumulator and the its the PassThru value we need to
9950     // bypass the negate. Conveniently Upper should be the same thing in this
9951     // case.
9952     if (NegAcc && PTIdx == 2)
9953       PassThru = CGF.Builder.CreateExtractElement(Upper, (uint64_t)0);
9954 
9955     Res = EmitX86ScalarSelect(CGF, Ops[3], Res, PassThru);
9956   }
9957   return CGF.Builder.CreateInsertElement(Upper, Res, (uint64_t)0);
9958 }
9959 
9960 static Value *EmitX86Muldq(CodeGenFunction &CGF, bool IsSigned,
9961                            ArrayRef<Value *> Ops) {
9962   llvm::Type *Ty = Ops[0]->getType();
9963   // Arguments have a vXi32 type so cast to vXi64.
9964   Ty = llvm::VectorType::get(CGF.Int64Ty,
9965                              Ty->getPrimitiveSizeInBits() / 64);
9966   Value *LHS = CGF.Builder.CreateBitCast(Ops[0], Ty);
9967   Value *RHS = CGF.Builder.CreateBitCast(Ops[1], Ty);
9968 
9969   if (IsSigned) {
9970     // Shift left then arithmetic shift right.
9971     Constant *ShiftAmt = ConstantInt::get(Ty, 32);
9972     LHS = CGF.Builder.CreateShl(LHS, ShiftAmt);
9973     LHS = CGF.Builder.CreateAShr(LHS, ShiftAmt);
9974     RHS = CGF.Builder.CreateShl(RHS, ShiftAmt);
9975     RHS = CGF.Builder.CreateAShr(RHS, ShiftAmt);
9976   } else {
9977     // Clear the upper bits.
9978     Constant *Mask = ConstantInt::get(Ty, 0xffffffff);
9979     LHS = CGF.Builder.CreateAnd(LHS, Mask);
9980     RHS = CGF.Builder.CreateAnd(RHS, Mask);
9981   }
9982 
9983   return CGF.Builder.CreateMul(LHS, RHS);
9984 }
9985 
9986 // Emit a masked pternlog intrinsic. This only exists because the header has to
9987 // use a macro and we aren't able to pass the input argument to a pternlog
9988 // builtin and a select builtin without evaluating it twice.
9989 static Value *EmitX86Ternlog(CodeGenFunction &CGF, bool ZeroMask,
9990                              ArrayRef<Value *> Ops) {
9991   llvm::Type *Ty = Ops[0]->getType();
9992 
9993   unsigned VecWidth = Ty->getPrimitiveSizeInBits();
9994   unsigned EltWidth = Ty->getScalarSizeInBits();
9995   Intrinsic::ID IID;
9996   if (VecWidth == 128 && EltWidth == 32)
9997     IID = Intrinsic::x86_avx512_pternlog_d_128;
9998   else if (VecWidth == 256 && EltWidth == 32)
9999     IID = Intrinsic::x86_avx512_pternlog_d_256;
10000   else if (VecWidth == 512 && EltWidth == 32)
10001     IID = Intrinsic::x86_avx512_pternlog_d_512;
10002   else if (VecWidth == 128 && EltWidth == 64)
10003     IID = Intrinsic::x86_avx512_pternlog_q_128;
10004   else if (VecWidth == 256 && EltWidth == 64)
10005     IID = Intrinsic::x86_avx512_pternlog_q_256;
10006   else if (VecWidth == 512 && EltWidth == 64)
10007     IID = Intrinsic::x86_avx512_pternlog_q_512;
10008   else
10009     llvm_unreachable("Unexpected intrinsic");
10010 
10011   Value *Ternlog = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID),
10012                                           Ops.drop_back());
10013   Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty) : Ops[0];
10014   return EmitX86Select(CGF, Ops[4], Ternlog, PassThru);
10015 }
10016 
10017 static Value *EmitX86SExtMask(CodeGenFunction &CGF, Value *Op,
10018                               llvm::Type *DstTy) {
10019   unsigned NumberOfElements = DstTy->getVectorNumElements();
10020   Value *Mask = getMaskVecValue(CGF, Op, NumberOfElements);
10021   return CGF.Builder.CreateSExt(Mask, DstTy, "vpmovm2");
10022 }
10023 
10024 // Emit addition or subtraction with signed/unsigned saturation.
10025 static Value *EmitX86AddSubSatExpr(CodeGenFunction &CGF,
10026                                    ArrayRef<Value *> Ops, bool IsSigned,
10027                                    bool IsAddition) {
10028   Intrinsic::ID IID =
10029       IsSigned ? (IsAddition ? Intrinsic::sadd_sat : Intrinsic::ssub_sat)
10030                : (IsAddition ? Intrinsic::uadd_sat : Intrinsic::usub_sat);
10031   llvm::Function *F = CGF.CGM.getIntrinsic(IID, Ops[0]->getType());
10032   return CGF.Builder.CreateCall(F, {Ops[0], Ops[1]});
10033 }
10034 
10035 Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) {
10036   const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts();
10037   StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString();
10038   return EmitX86CpuIs(CPUStr);
10039 }
10040 
10041 // Convert a BF16 to a float.
10042 static Value *EmitX86CvtBF16ToFloatExpr(CodeGenFunction &CGF,
10043                                         const CallExpr *E,
10044                                         ArrayRef<Value *> Ops) {
10045   llvm::Type *Int32Ty = CGF.Builder.getInt32Ty();
10046   Value *ZeroExt = CGF.Builder.CreateZExt(Ops[0], Int32Ty);
10047   Value *Shl = CGF.Builder.CreateShl(ZeroExt, 16);
10048   llvm::Type *ResultType = CGF.ConvertType(E->getType());
10049   Value *BitCast = CGF.Builder.CreateBitCast(Shl, ResultType);
10050   return BitCast;
10051 }
10052 
10053 Value *CodeGenFunction::EmitX86CpuIs(StringRef CPUStr) {
10054 
10055   llvm::Type *Int32Ty = Builder.getInt32Ty();
10056 
10057   // Matching the struct layout from the compiler-rt/libgcc structure that is
10058   // filled in:
10059   // unsigned int __cpu_vendor;
10060   // unsigned int __cpu_type;
10061   // unsigned int __cpu_subtype;
10062   // unsigned int __cpu_features[1];
10063   llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty,
10064                                           llvm::ArrayType::get(Int32Ty, 1));
10065 
10066   // Grab the global __cpu_model.
10067   llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model");
10068   cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true);
10069 
10070   // Calculate the index needed to access the correct field based on the
10071   // range. Also adjust the expected value.
10072   unsigned Index;
10073   unsigned Value;
10074   std::tie(Index, Value) = StringSwitch<std::pair<unsigned, unsigned>>(CPUStr)
10075 #define X86_VENDOR(ENUM, STRING)                                               \
10076   .Case(STRING, {0u, static_cast<unsigned>(llvm::X86::ENUM)})
10077 #define X86_CPU_TYPE_COMPAT_WITH_ALIAS(ARCHNAME, ENUM, STR, ALIAS)             \
10078   .Cases(STR, ALIAS, {1u, static_cast<unsigned>(llvm::X86::ENUM)})
10079 #define X86_CPU_TYPE_COMPAT(ARCHNAME, ENUM, STR)                               \
10080   .Case(STR, {1u, static_cast<unsigned>(llvm::X86::ENUM)})
10081 #define X86_CPU_SUBTYPE_COMPAT(ARCHNAME, ENUM, STR)                            \
10082   .Case(STR, {2u, static_cast<unsigned>(llvm::X86::ENUM)})
10083 #include "llvm/Support/X86TargetParser.def"
10084                                .Default({0, 0});
10085   assert(Value != 0 && "Invalid CPUStr passed to CpuIs");
10086 
10087   // Grab the appropriate field from __cpu_model.
10088   llvm::Value *Idxs[] = {ConstantInt::get(Int32Ty, 0),
10089                          ConstantInt::get(Int32Ty, Index)};
10090   llvm::Value *CpuValue = Builder.CreateGEP(STy, CpuModel, Idxs);
10091   CpuValue = Builder.CreateAlignedLoad(CpuValue, CharUnits::fromQuantity(4));
10092 
10093   // Check the value of the field against the requested value.
10094   return Builder.CreateICmpEQ(CpuValue,
10095                                   llvm::ConstantInt::get(Int32Ty, Value));
10096 }
10097 
10098 Value *CodeGenFunction::EmitX86CpuSupports(const CallExpr *E) {
10099   const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts();
10100   StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString();
10101   return EmitX86CpuSupports(FeatureStr);
10102 }
10103 
10104 uint64_t
10105 CodeGenFunction::GetX86CpuSupportsMask(ArrayRef<StringRef> FeatureStrs) {
10106   // Processor features and mapping to processor feature value.
10107   uint64_t FeaturesMask = 0;
10108   for (const StringRef &FeatureStr : FeatureStrs) {
10109     unsigned Feature =
10110         StringSwitch<unsigned>(FeatureStr)
10111 #define X86_FEATURE_COMPAT(VAL, ENUM, STR) .Case(STR, VAL)
10112 #include "llvm/Support/X86TargetParser.def"
10113         ;
10114     FeaturesMask |= (1ULL << Feature);
10115   }
10116   return FeaturesMask;
10117 }
10118 
10119 Value *CodeGenFunction::EmitX86CpuSupports(ArrayRef<StringRef> FeatureStrs) {
10120   return EmitX86CpuSupports(GetX86CpuSupportsMask(FeatureStrs));
10121 }
10122 
10123 llvm::Value *CodeGenFunction::EmitX86CpuSupports(uint64_t FeaturesMask) {
10124   uint32_t Features1 = Lo_32(FeaturesMask);
10125   uint32_t Features2 = Hi_32(FeaturesMask);
10126 
10127   Value *Result = Builder.getTrue();
10128 
10129   if (Features1 != 0) {
10130     // Matching the struct layout from the compiler-rt/libgcc structure that is
10131     // filled in:
10132     // unsigned int __cpu_vendor;
10133     // unsigned int __cpu_type;
10134     // unsigned int __cpu_subtype;
10135     // unsigned int __cpu_features[1];
10136     llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty,
10137                                             llvm::ArrayType::get(Int32Ty, 1));
10138 
10139     // Grab the global __cpu_model.
10140     llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model");
10141     cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true);
10142 
10143     // Grab the first (0th) element from the field __cpu_features off of the
10144     // global in the struct STy.
10145     Value *Idxs[] = {Builder.getInt32(0), Builder.getInt32(3),
10146                      Builder.getInt32(0)};
10147     Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs);
10148     Value *Features =
10149         Builder.CreateAlignedLoad(CpuFeatures, CharUnits::fromQuantity(4));
10150 
10151     // Check the value of the bit corresponding to the feature requested.
10152     Value *Mask = Builder.getInt32(Features1);
10153     Value *Bitset = Builder.CreateAnd(Features, Mask);
10154     Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
10155     Result = Builder.CreateAnd(Result, Cmp);
10156   }
10157 
10158   if (Features2 != 0) {
10159     llvm::Constant *CpuFeatures2 = CGM.CreateRuntimeVariable(Int32Ty,
10160                                                              "__cpu_features2");
10161     cast<llvm::GlobalValue>(CpuFeatures2)->setDSOLocal(true);
10162 
10163     Value *Features =
10164         Builder.CreateAlignedLoad(CpuFeatures2, CharUnits::fromQuantity(4));
10165 
10166     // Check the value of the bit corresponding to the feature requested.
10167     Value *Mask = Builder.getInt32(Features2);
10168     Value *Bitset = Builder.CreateAnd(Features, Mask);
10169     Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
10170     Result = Builder.CreateAnd(Result, Cmp);
10171   }
10172 
10173   return Result;
10174 }
10175 
10176 Value *CodeGenFunction::EmitX86CpuInit() {
10177   llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy,
10178                                                     /*Variadic*/ false);
10179   llvm::FunctionCallee Func =
10180       CGM.CreateRuntimeFunction(FTy, "__cpu_indicator_init");
10181   cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true);
10182   cast<llvm::GlobalValue>(Func.getCallee())
10183       ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass);
10184   return Builder.CreateCall(Func);
10185 }
10186 
10187 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
10188                                            const CallExpr *E) {
10189   if (BuiltinID == X86::BI__builtin_cpu_is)
10190     return EmitX86CpuIs(E);
10191   if (BuiltinID == X86::BI__builtin_cpu_supports)
10192     return EmitX86CpuSupports(E);
10193   if (BuiltinID == X86::BI__builtin_cpu_init)
10194     return EmitX86CpuInit();
10195 
10196   SmallVector<Value*, 4> Ops;
10197 
10198   // Find out if any arguments are required to be integer constant expressions.
10199   unsigned ICEArguments = 0;
10200   ASTContext::GetBuiltinTypeError Error;
10201   getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
10202   assert(Error == ASTContext::GE_None && "Should not codegen an error");
10203 
10204   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
10205     // If this is a normal argument, just emit it as a scalar.
10206     if ((ICEArguments & (1 << i)) == 0) {
10207       Ops.push_back(EmitScalarExpr(E->getArg(i)));
10208       continue;
10209     }
10210 
10211     // If this is required to be a constant, constant fold it so that we know
10212     // that the generated intrinsic gets a ConstantInt.
10213     llvm::APSInt Result;
10214     bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
10215     assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
10216     Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
10217   }
10218 
10219   // These exist so that the builtin that takes an immediate can be bounds
10220   // checked by clang to avoid passing bad immediates to the backend. Since
10221   // AVX has a larger immediate than SSE we would need separate builtins to
10222   // do the different bounds checking. Rather than create a clang specific
10223   // SSE only builtin, this implements eight separate builtins to match gcc
10224   // implementation.
10225   auto getCmpIntrinsicCall = [this, &Ops](Intrinsic::ID ID, unsigned Imm) {
10226     Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm));
10227     llvm::Function *F = CGM.getIntrinsic(ID);
10228     return Builder.CreateCall(F, Ops);
10229   };
10230 
10231   // For the vector forms of FP comparisons, translate the builtins directly to
10232   // IR.
10233   // TODO: The builtins could be removed if the SSE header files used vector
10234   // extension comparisons directly (vector ordered/unordered may need
10235   // additional support via __builtin_isnan()).
10236   auto getVectorFCmpIR = [this, &Ops](CmpInst::Predicate Pred) {
10237     Value *Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]);
10238     llvm::VectorType *FPVecTy = cast<llvm::VectorType>(Ops[0]->getType());
10239     llvm::VectorType *IntVecTy = llvm::VectorType::getInteger(FPVecTy);
10240     Value *Sext = Builder.CreateSExt(Cmp, IntVecTy);
10241     return Builder.CreateBitCast(Sext, FPVecTy);
10242   };
10243 
10244   switch (BuiltinID) {
10245   default: return nullptr;
10246   case X86::BI_mm_prefetch: {
10247     Value *Address = Ops[0];
10248     ConstantInt *C = cast<ConstantInt>(Ops[1]);
10249     Value *RW = ConstantInt::get(Int32Ty, (C->getZExtValue() >> 2) & 0x1);
10250     Value *Locality = ConstantInt::get(Int32Ty, C->getZExtValue() & 0x3);
10251     Value *Data = ConstantInt::get(Int32Ty, 1);
10252     Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
10253     return Builder.CreateCall(F, {Address, RW, Locality, Data});
10254   }
10255   case X86::BI_mm_clflush: {
10256     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_clflush),
10257                               Ops[0]);
10258   }
10259   case X86::BI_mm_lfence: {
10260     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_lfence));
10261   }
10262   case X86::BI_mm_mfence: {
10263     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_mfence));
10264   }
10265   case X86::BI_mm_sfence: {
10266     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_sfence));
10267   }
10268   case X86::BI_mm_pause: {
10269     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_pause));
10270   }
10271   case X86::BI__rdtsc: {
10272     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtsc));
10273   }
10274   case X86::BI__builtin_ia32_rdtscp: {
10275     Value *Call = Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtscp));
10276     Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1),
10277                                       Ops[0]);
10278     return Builder.CreateExtractValue(Call, 0);
10279   }
10280   case X86::BI__builtin_ia32_lzcnt_u16:
10281   case X86::BI__builtin_ia32_lzcnt_u32:
10282   case X86::BI__builtin_ia32_lzcnt_u64: {
10283     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
10284     return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
10285   }
10286   case X86::BI__builtin_ia32_tzcnt_u16:
10287   case X86::BI__builtin_ia32_tzcnt_u32:
10288   case X86::BI__builtin_ia32_tzcnt_u64: {
10289     Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType());
10290     return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
10291   }
10292   case X86::BI__builtin_ia32_undef128:
10293   case X86::BI__builtin_ia32_undef256:
10294   case X86::BI__builtin_ia32_undef512:
10295     // The x86 definition of "undef" is not the same as the LLVM definition
10296     // (PR32176). We leave optimizing away an unnecessary zero constant to the
10297     // IR optimizer and backend.
10298     // TODO: If we had a "freeze" IR instruction to generate a fixed undef
10299     // value, we should use that here instead of a zero.
10300     return llvm::Constant::getNullValue(ConvertType(E->getType()));
10301   case X86::BI__builtin_ia32_vec_init_v8qi:
10302   case X86::BI__builtin_ia32_vec_init_v4hi:
10303   case X86::BI__builtin_ia32_vec_init_v2si:
10304     return Builder.CreateBitCast(BuildVector(Ops),
10305                                  llvm::Type::getX86_MMXTy(getLLVMContext()));
10306   case X86::BI__builtin_ia32_vec_ext_v2si:
10307   case X86::BI__builtin_ia32_vec_ext_v16qi:
10308   case X86::BI__builtin_ia32_vec_ext_v8hi:
10309   case X86::BI__builtin_ia32_vec_ext_v4si:
10310   case X86::BI__builtin_ia32_vec_ext_v4sf:
10311   case X86::BI__builtin_ia32_vec_ext_v2di:
10312   case X86::BI__builtin_ia32_vec_ext_v32qi:
10313   case X86::BI__builtin_ia32_vec_ext_v16hi:
10314   case X86::BI__builtin_ia32_vec_ext_v8si:
10315   case X86::BI__builtin_ia32_vec_ext_v4di: {
10316     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
10317     uint64_t Index = cast<ConstantInt>(Ops[1])->getZExtValue();
10318     Index &= NumElts - 1;
10319     // These builtins exist so we can ensure the index is an ICE and in range.
10320     // Otherwise we could just do this in the header file.
10321     return Builder.CreateExtractElement(Ops[0], Index);
10322   }
10323   case X86::BI__builtin_ia32_vec_set_v16qi:
10324   case X86::BI__builtin_ia32_vec_set_v8hi:
10325   case X86::BI__builtin_ia32_vec_set_v4si:
10326   case X86::BI__builtin_ia32_vec_set_v2di:
10327   case X86::BI__builtin_ia32_vec_set_v32qi:
10328   case X86::BI__builtin_ia32_vec_set_v16hi:
10329   case X86::BI__builtin_ia32_vec_set_v8si:
10330   case X86::BI__builtin_ia32_vec_set_v4di: {
10331     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
10332     unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue();
10333     Index &= NumElts - 1;
10334     // These builtins exist so we can ensure the index is an ICE and in range.
10335     // Otherwise we could just do this in the header file.
10336     return Builder.CreateInsertElement(Ops[0], Ops[1], Index);
10337   }
10338   case X86::BI_mm_setcsr:
10339   case X86::BI__builtin_ia32_ldmxcsr: {
10340     Address Tmp = CreateMemTemp(E->getArg(0)->getType());
10341     Builder.CreateStore(Ops[0], Tmp);
10342     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
10343                           Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy));
10344   }
10345   case X86::BI_mm_getcsr:
10346   case X86::BI__builtin_ia32_stmxcsr: {
10347     Address Tmp = CreateMemTemp(E->getType());
10348     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
10349                        Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy));
10350     return Builder.CreateLoad(Tmp, "stmxcsr");
10351   }
10352   case X86::BI__builtin_ia32_xsave:
10353   case X86::BI__builtin_ia32_xsave64:
10354   case X86::BI__builtin_ia32_xrstor:
10355   case X86::BI__builtin_ia32_xrstor64:
10356   case X86::BI__builtin_ia32_xsaveopt:
10357   case X86::BI__builtin_ia32_xsaveopt64:
10358   case X86::BI__builtin_ia32_xrstors:
10359   case X86::BI__builtin_ia32_xrstors64:
10360   case X86::BI__builtin_ia32_xsavec:
10361   case X86::BI__builtin_ia32_xsavec64:
10362   case X86::BI__builtin_ia32_xsaves:
10363   case X86::BI__builtin_ia32_xsaves64:
10364   case X86::BI__builtin_ia32_xsetbv:
10365   case X86::BI_xsetbv: {
10366     Intrinsic::ID ID;
10367 #define INTRINSIC_X86_XSAVE_ID(NAME) \
10368     case X86::BI__builtin_ia32_##NAME: \
10369       ID = Intrinsic::x86_##NAME; \
10370       break
10371     switch (BuiltinID) {
10372     default: llvm_unreachable("Unsupported intrinsic!");
10373     INTRINSIC_X86_XSAVE_ID(xsave);
10374     INTRINSIC_X86_XSAVE_ID(xsave64);
10375     INTRINSIC_X86_XSAVE_ID(xrstor);
10376     INTRINSIC_X86_XSAVE_ID(xrstor64);
10377     INTRINSIC_X86_XSAVE_ID(xsaveopt);
10378     INTRINSIC_X86_XSAVE_ID(xsaveopt64);
10379     INTRINSIC_X86_XSAVE_ID(xrstors);
10380     INTRINSIC_X86_XSAVE_ID(xrstors64);
10381     INTRINSIC_X86_XSAVE_ID(xsavec);
10382     INTRINSIC_X86_XSAVE_ID(xsavec64);
10383     INTRINSIC_X86_XSAVE_ID(xsaves);
10384     INTRINSIC_X86_XSAVE_ID(xsaves64);
10385     INTRINSIC_X86_XSAVE_ID(xsetbv);
10386     case X86::BI_xsetbv:
10387       ID = Intrinsic::x86_xsetbv;
10388       break;
10389     }
10390 #undef INTRINSIC_X86_XSAVE_ID
10391     Value *Mhi = Builder.CreateTrunc(
10392       Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty);
10393     Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty);
10394     Ops[1] = Mhi;
10395     Ops.push_back(Mlo);
10396     return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
10397   }
10398   case X86::BI__builtin_ia32_xgetbv:
10399   case X86::BI_xgetbv:
10400     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_xgetbv), Ops);
10401   case X86::BI__builtin_ia32_storedqudi128_mask:
10402   case X86::BI__builtin_ia32_storedqusi128_mask:
10403   case X86::BI__builtin_ia32_storedquhi128_mask:
10404   case X86::BI__builtin_ia32_storedquqi128_mask:
10405   case X86::BI__builtin_ia32_storeupd128_mask:
10406   case X86::BI__builtin_ia32_storeups128_mask:
10407   case X86::BI__builtin_ia32_storedqudi256_mask:
10408   case X86::BI__builtin_ia32_storedqusi256_mask:
10409   case X86::BI__builtin_ia32_storedquhi256_mask:
10410   case X86::BI__builtin_ia32_storedquqi256_mask:
10411   case X86::BI__builtin_ia32_storeupd256_mask:
10412   case X86::BI__builtin_ia32_storeups256_mask:
10413   case X86::BI__builtin_ia32_storedqudi512_mask:
10414   case X86::BI__builtin_ia32_storedqusi512_mask:
10415   case X86::BI__builtin_ia32_storedquhi512_mask:
10416   case X86::BI__builtin_ia32_storedquqi512_mask:
10417   case X86::BI__builtin_ia32_storeupd512_mask:
10418   case X86::BI__builtin_ia32_storeups512_mask:
10419     return EmitX86MaskedStore(*this, Ops, 1);
10420 
10421   case X86::BI__builtin_ia32_storess128_mask:
10422   case X86::BI__builtin_ia32_storesd128_mask: {
10423     return EmitX86MaskedStore(*this, Ops, 1);
10424   }
10425   case X86::BI__builtin_ia32_vpopcntb_128:
10426   case X86::BI__builtin_ia32_vpopcntd_128:
10427   case X86::BI__builtin_ia32_vpopcntq_128:
10428   case X86::BI__builtin_ia32_vpopcntw_128:
10429   case X86::BI__builtin_ia32_vpopcntb_256:
10430   case X86::BI__builtin_ia32_vpopcntd_256:
10431   case X86::BI__builtin_ia32_vpopcntq_256:
10432   case X86::BI__builtin_ia32_vpopcntw_256:
10433   case X86::BI__builtin_ia32_vpopcntb_512:
10434   case X86::BI__builtin_ia32_vpopcntd_512:
10435   case X86::BI__builtin_ia32_vpopcntq_512:
10436   case X86::BI__builtin_ia32_vpopcntw_512: {
10437     llvm::Type *ResultType = ConvertType(E->getType());
10438     llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
10439     return Builder.CreateCall(F, Ops);
10440   }
10441   case X86::BI__builtin_ia32_cvtmask2b128:
10442   case X86::BI__builtin_ia32_cvtmask2b256:
10443   case X86::BI__builtin_ia32_cvtmask2b512:
10444   case X86::BI__builtin_ia32_cvtmask2w128:
10445   case X86::BI__builtin_ia32_cvtmask2w256:
10446   case X86::BI__builtin_ia32_cvtmask2w512:
10447   case X86::BI__builtin_ia32_cvtmask2d128:
10448   case X86::BI__builtin_ia32_cvtmask2d256:
10449   case X86::BI__builtin_ia32_cvtmask2d512:
10450   case X86::BI__builtin_ia32_cvtmask2q128:
10451   case X86::BI__builtin_ia32_cvtmask2q256:
10452   case X86::BI__builtin_ia32_cvtmask2q512:
10453     return EmitX86SExtMask(*this, Ops[0], ConvertType(E->getType()));
10454 
10455   case X86::BI__builtin_ia32_cvtb2mask128:
10456   case X86::BI__builtin_ia32_cvtb2mask256:
10457   case X86::BI__builtin_ia32_cvtb2mask512:
10458   case X86::BI__builtin_ia32_cvtw2mask128:
10459   case X86::BI__builtin_ia32_cvtw2mask256:
10460   case X86::BI__builtin_ia32_cvtw2mask512:
10461   case X86::BI__builtin_ia32_cvtd2mask128:
10462   case X86::BI__builtin_ia32_cvtd2mask256:
10463   case X86::BI__builtin_ia32_cvtd2mask512:
10464   case X86::BI__builtin_ia32_cvtq2mask128:
10465   case X86::BI__builtin_ia32_cvtq2mask256:
10466   case X86::BI__builtin_ia32_cvtq2mask512:
10467     return EmitX86ConvertToMask(*this, Ops[0]);
10468 
10469   case X86::BI__builtin_ia32_cvtdq2ps512_mask:
10470   case X86::BI__builtin_ia32_cvtqq2ps512_mask:
10471   case X86::BI__builtin_ia32_cvtqq2pd512_mask:
10472     return EmitX86ConvertIntToFp(*this, Ops, /*IsSigned*/true);
10473   case X86::BI__builtin_ia32_cvtudq2ps512_mask:
10474   case X86::BI__builtin_ia32_cvtuqq2ps512_mask:
10475   case X86::BI__builtin_ia32_cvtuqq2pd512_mask:
10476     return EmitX86ConvertIntToFp(*this, Ops, /*IsSigned*/false);
10477 
10478   case X86::BI__builtin_ia32_vfmaddss3:
10479   case X86::BI__builtin_ia32_vfmaddsd3:
10480   case X86::BI__builtin_ia32_vfmaddss3_mask:
10481   case X86::BI__builtin_ia32_vfmaddsd3_mask:
10482     return EmitScalarFMAExpr(*this, Ops, Ops[0]);
10483   case X86::BI__builtin_ia32_vfmaddss:
10484   case X86::BI__builtin_ia32_vfmaddsd:
10485     return EmitScalarFMAExpr(*this, Ops,
10486                              Constant::getNullValue(Ops[0]->getType()));
10487   case X86::BI__builtin_ia32_vfmaddss3_maskz:
10488   case X86::BI__builtin_ia32_vfmaddsd3_maskz:
10489     return EmitScalarFMAExpr(*this, Ops, Ops[0], /*ZeroMask*/true);
10490   case X86::BI__builtin_ia32_vfmaddss3_mask3:
10491   case X86::BI__builtin_ia32_vfmaddsd3_mask3:
10492     return EmitScalarFMAExpr(*this, Ops, Ops[2], /*ZeroMask*/false, 2);
10493   case X86::BI__builtin_ia32_vfmsubss3_mask3:
10494   case X86::BI__builtin_ia32_vfmsubsd3_mask3:
10495     return EmitScalarFMAExpr(*this, Ops, Ops[2], /*ZeroMask*/false, 2,
10496                              /*NegAcc*/true);
10497   case X86::BI__builtin_ia32_vfmaddps:
10498   case X86::BI__builtin_ia32_vfmaddpd:
10499   case X86::BI__builtin_ia32_vfmaddps256:
10500   case X86::BI__builtin_ia32_vfmaddpd256:
10501   case X86::BI__builtin_ia32_vfmaddps512_mask:
10502   case X86::BI__builtin_ia32_vfmaddps512_maskz:
10503   case X86::BI__builtin_ia32_vfmaddps512_mask3:
10504   case X86::BI__builtin_ia32_vfmsubps512_mask3:
10505   case X86::BI__builtin_ia32_vfmaddpd512_mask:
10506   case X86::BI__builtin_ia32_vfmaddpd512_maskz:
10507   case X86::BI__builtin_ia32_vfmaddpd512_mask3:
10508   case X86::BI__builtin_ia32_vfmsubpd512_mask3:
10509     return EmitX86FMAExpr(*this, Ops, BuiltinID, /*IsAddSub*/false);
10510   case X86::BI__builtin_ia32_vfmaddsubps:
10511   case X86::BI__builtin_ia32_vfmaddsubpd:
10512   case X86::BI__builtin_ia32_vfmaddsubps256:
10513   case X86::BI__builtin_ia32_vfmaddsubpd256:
10514   case X86::BI__builtin_ia32_vfmaddsubps512_mask:
10515   case X86::BI__builtin_ia32_vfmaddsubps512_maskz:
10516   case X86::BI__builtin_ia32_vfmaddsubps512_mask3:
10517   case X86::BI__builtin_ia32_vfmsubaddps512_mask3:
10518   case X86::BI__builtin_ia32_vfmaddsubpd512_mask:
10519   case X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
10520   case X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
10521   case X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
10522     return EmitX86FMAExpr(*this, Ops, BuiltinID, /*IsAddSub*/true);
10523 
10524   case X86::BI__builtin_ia32_movdqa32store128_mask:
10525   case X86::BI__builtin_ia32_movdqa64store128_mask:
10526   case X86::BI__builtin_ia32_storeaps128_mask:
10527   case X86::BI__builtin_ia32_storeapd128_mask:
10528   case X86::BI__builtin_ia32_movdqa32store256_mask:
10529   case X86::BI__builtin_ia32_movdqa64store256_mask:
10530   case X86::BI__builtin_ia32_storeaps256_mask:
10531   case X86::BI__builtin_ia32_storeapd256_mask:
10532   case X86::BI__builtin_ia32_movdqa32store512_mask:
10533   case X86::BI__builtin_ia32_movdqa64store512_mask:
10534   case X86::BI__builtin_ia32_storeaps512_mask:
10535   case X86::BI__builtin_ia32_storeapd512_mask: {
10536     unsigned Align =
10537       getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity();
10538     return EmitX86MaskedStore(*this, Ops, Align);
10539   }
10540   case X86::BI__builtin_ia32_loadups128_mask:
10541   case X86::BI__builtin_ia32_loadups256_mask:
10542   case X86::BI__builtin_ia32_loadups512_mask:
10543   case X86::BI__builtin_ia32_loadupd128_mask:
10544   case X86::BI__builtin_ia32_loadupd256_mask:
10545   case X86::BI__builtin_ia32_loadupd512_mask:
10546   case X86::BI__builtin_ia32_loaddquqi128_mask:
10547   case X86::BI__builtin_ia32_loaddquqi256_mask:
10548   case X86::BI__builtin_ia32_loaddquqi512_mask:
10549   case X86::BI__builtin_ia32_loaddquhi128_mask:
10550   case X86::BI__builtin_ia32_loaddquhi256_mask:
10551   case X86::BI__builtin_ia32_loaddquhi512_mask:
10552   case X86::BI__builtin_ia32_loaddqusi128_mask:
10553   case X86::BI__builtin_ia32_loaddqusi256_mask:
10554   case X86::BI__builtin_ia32_loaddqusi512_mask:
10555   case X86::BI__builtin_ia32_loaddqudi128_mask:
10556   case X86::BI__builtin_ia32_loaddqudi256_mask:
10557   case X86::BI__builtin_ia32_loaddqudi512_mask:
10558     return EmitX86MaskedLoad(*this, Ops, 1);
10559 
10560   case X86::BI__builtin_ia32_loadss128_mask:
10561   case X86::BI__builtin_ia32_loadsd128_mask:
10562     return EmitX86MaskedLoad(*this, Ops, 1);
10563 
10564   case X86::BI__builtin_ia32_loadaps128_mask:
10565   case X86::BI__builtin_ia32_loadaps256_mask:
10566   case X86::BI__builtin_ia32_loadaps512_mask:
10567   case X86::BI__builtin_ia32_loadapd128_mask:
10568   case X86::BI__builtin_ia32_loadapd256_mask:
10569   case X86::BI__builtin_ia32_loadapd512_mask:
10570   case X86::BI__builtin_ia32_movdqa32load128_mask:
10571   case X86::BI__builtin_ia32_movdqa32load256_mask:
10572   case X86::BI__builtin_ia32_movdqa32load512_mask:
10573   case X86::BI__builtin_ia32_movdqa64load128_mask:
10574   case X86::BI__builtin_ia32_movdqa64load256_mask:
10575   case X86::BI__builtin_ia32_movdqa64load512_mask: {
10576     unsigned Align =
10577       getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity();
10578     return EmitX86MaskedLoad(*this, Ops, Align);
10579   }
10580 
10581   case X86::BI__builtin_ia32_expandloaddf128_mask:
10582   case X86::BI__builtin_ia32_expandloaddf256_mask:
10583   case X86::BI__builtin_ia32_expandloaddf512_mask:
10584   case X86::BI__builtin_ia32_expandloadsf128_mask:
10585   case X86::BI__builtin_ia32_expandloadsf256_mask:
10586   case X86::BI__builtin_ia32_expandloadsf512_mask:
10587   case X86::BI__builtin_ia32_expandloaddi128_mask:
10588   case X86::BI__builtin_ia32_expandloaddi256_mask:
10589   case X86::BI__builtin_ia32_expandloaddi512_mask:
10590   case X86::BI__builtin_ia32_expandloadsi128_mask:
10591   case X86::BI__builtin_ia32_expandloadsi256_mask:
10592   case X86::BI__builtin_ia32_expandloadsi512_mask:
10593   case X86::BI__builtin_ia32_expandloadhi128_mask:
10594   case X86::BI__builtin_ia32_expandloadhi256_mask:
10595   case X86::BI__builtin_ia32_expandloadhi512_mask:
10596   case X86::BI__builtin_ia32_expandloadqi128_mask:
10597   case X86::BI__builtin_ia32_expandloadqi256_mask:
10598   case X86::BI__builtin_ia32_expandloadqi512_mask:
10599     return EmitX86ExpandLoad(*this, Ops);
10600 
10601   case X86::BI__builtin_ia32_compressstoredf128_mask:
10602   case X86::BI__builtin_ia32_compressstoredf256_mask:
10603   case X86::BI__builtin_ia32_compressstoredf512_mask:
10604   case X86::BI__builtin_ia32_compressstoresf128_mask:
10605   case X86::BI__builtin_ia32_compressstoresf256_mask:
10606   case X86::BI__builtin_ia32_compressstoresf512_mask:
10607   case X86::BI__builtin_ia32_compressstoredi128_mask:
10608   case X86::BI__builtin_ia32_compressstoredi256_mask:
10609   case X86::BI__builtin_ia32_compressstoredi512_mask:
10610   case X86::BI__builtin_ia32_compressstoresi128_mask:
10611   case X86::BI__builtin_ia32_compressstoresi256_mask:
10612   case X86::BI__builtin_ia32_compressstoresi512_mask:
10613   case X86::BI__builtin_ia32_compressstorehi128_mask:
10614   case X86::BI__builtin_ia32_compressstorehi256_mask:
10615   case X86::BI__builtin_ia32_compressstorehi512_mask:
10616   case X86::BI__builtin_ia32_compressstoreqi128_mask:
10617   case X86::BI__builtin_ia32_compressstoreqi256_mask:
10618   case X86::BI__builtin_ia32_compressstoreqi512_mask:
10619     return EmitX86CompressStore(*this, Ops);
10620 
10621   case X86::BI__builtin_ia32_expanddf128_mask:
10622   case X86::BI__builtin_ia32_expanddf256_mask:
10623   case X86::BI__builtin_ia32_expanddf512_mask:
10624   case X86::BI__builtin_ia32_expandsf128_mask:
10625   case X86::BI__builtin_ia32_expandsf256_mask:
10626   case X86::BI__builtin_ia32_expandsf512_mask:
10627   case X86::BI__builtin_ia32_expanddi128_mask:
10628   case X86::BI__builtin_ia32_expanddi256_mask:
10629   case X86::BI__builtin_ia32_expanddi512_mask:
10630   case X86::BI__builtin_ia32_expandsi128_mask:
10631   case X86::BI__builtin_ia32_expandsi256_mask:
10632   case X86::BI__builtin_ia32_expandsi512_mask:
10633   case X86::BI__builtin_ia32_expandhi128_mask:
10634   case X86::BI__builtin_ia32_expandhi256_mask:
10635   case X86::BI__builtin_ia32_expandhi512_mask:
10636   case X86::BI__builtin_ia32_expandqi128_mask:
10637   case X86::BI__builtin_ia32_expandqi256_mask:
10638   case X86::BI__builtin_ia32_expandqi512_mask:
10639     return EmitX86CompressExpand(*this, Ops, /*IsCompress*/false);
10640 
10641   case X86::BI__builtin_ia32_compressdf128_mask:
10642   case X86::BI__builtin_ia32_compressdf256_mask:
10643   case X86::BI__builtin_ia32_compressdf512_mask:
10644   case X86::BI__builtin_ia32_compresssf128_mask:
10645   case X86::BI__builtin_ia32_compresssf256_mask:
10646   case X86::BI__builtin_ia32_compresssf512_mask:
10647   case X86::BI__builtin_ia32_compressdi128_mask:
10648   case X86::BI__builtin_ia32_compressdi256_mask:
10649   case X86::BI__builtin_ia32_compressdi512_mask:
10650   case X86::BI__builtin_ia32_compresssi128_mask:
10651   case X86::BI__builtin_ia32_compresssi256_mask:
10652   case X86::BI__builtin_ia32_compresssi512_mask:
10653   case X86::BI__builtin_ia32_compresshi128_mask:
10654   case X86::BI__builtin_ia32_compresshi256_mask:
10655   case X86::BI__builtin_ia32_compresshi512_mask:
10656   case X86::BI__builtin_ia32_compressqi128_mask:
10657   case X86::BI__builtin_ia32_compressqi256_mask:
10658   case X86::BI__builtin_ia32_compressqi512_mask:
10659     return EmitX86CompressExpand(*this, Ops, /*IsCompress*/true);
10660 
10661   case X86::BI__builtin_ia32_gather3div2df:
10662   case X86::BI__builtin_ia32_gather3div2di:
10663   case X86::BI__builtin_ia32_gather3div4df:
10664   case X86::BI__builtin_ia32_gather3div4di:
10665   case X86::BI__builtin_ia32_gather3div4sf:
10666   case X86::BI__builtin_ia32_gather3div4si:
10667   case X86::BI__builtin_ia32_gather3div8sf:
10668   case X86::BI__builtin_ia32_gather3div8si:
10669   case X86::BI__builtin_ia32_gather3siv2df:
10670   case X86::BI__builtin_ia32_gather3siv2di:
10671   case X86::BI__builtin_ia32_gather3siv4df:
10672   case X86::BI__builtin_ia32_gather3siv4di:
10673   case X86::BI__builtin_ia32_gather3siv4sf:
10674   case X86::BI__builtin_ia32_gather3siv4si:
10675   case X86::BI__builtin_ia32_gather3siv8sf:
10676   case X86::BI__builtin_ia32_gather3siv8si:
10677   case X86::BI__builtin_ia32_gathersiv8df:
10678   case X86::BI__builtin_ia32_gathersiv16sf:
10679   case X86::BI__builtin_ia32_gatherdiv8df:
10680   case X86::BI__builtin_ia32_gatherdiv16sf:
10681   case X86::BI__builtin_ia32_gathersiv8di:
10682   case X86::BI__builtin_ia32_gathersiv16si:
10683   case X86::BI__builtin_ia32_gatherdiv8di:
10684   case X86::BI__builtin_ia32_gatherdiv16si: {
10685     Intrinsic::ID IID;
10686     switch (BuiltinID) {
10687     default: llvm_unreachable("Unexpected builtin");
10688     case X86::BI__builtin_ia32_gather3div2df:
10689       IID = Intrinsic::x86_avx512_mask_gather3div2_df;
10690       break;
10691     case X86::BI__builtin_ia32_gather3div2di:
10692       IID = Intrinsic::x86_avx512_mask_gather3div2_di;
10693       break;
10694     case X86::BI__builtin_ia32_gather3div4df:
10695       IID = Intrinsic::x86_avx512_mask_gather3div4_df;
10696       break;
10697     case X86::BI__builtin_ia32_gather3div4di:
10698       IID = Intrinsic::x86_avx512_mask_gather3div4_di;
10699       break;
10700     case X86::BI__builtin_ia32_gather3div4sf:
10701       IID = Intrinsic::x86_avx512_mask_gather3div4_sf;
10702       break;
10703     case X86::BI__builtin_ia32_gather3div4si:
10704       IID = Intrinsic::x86_avx512_mask_gather3div4_si;
10705       break;
10706     case X86::BI__builtin_ia32_gather3div8sf:
10707       IID = Intrinsic::x86_avx512_mask_gather3div8_sf;
10708       break;
10709     case X86::BI__builtin_ia32_gather3div8si:
10710       IID = Intrinsic::x86_avx512_mask_gather3div8_si;
10711       break;
10712     case X86::BI__builtin_ia32_gather3siv2df:
10713       IID = Intrinsic::x86_avx512_mask_gather3siv2_df;
10714       break;
10715     case X86::BI__builtin_ia32_gather3siv2di:
10716       IID = Intrinsic::x86_avx512_mask_gather3siv2_di;
10717       break;
10718     case X86::BI__builtin_ia32_gather3siv4df:
10719       IID = Intrinsic::x86_avx512_mask_gather3siv4_df;
10720       break;
10721     case X86::BI__builtin_ia32_gather3siv4di:
10722       IID = Intrinsic::x86_avx512_mask_gather3siv4_di;
10723       break;
10724     case X86::BI__builtin_ia32_gather3siv4sf:
10725       IID = Intrinsic::x86_avx512_mask_gather3siv4_sf;
10726       break;
10727     case X86::BI__builtin_ia32_gather3siv4si:
10728       IID = Intrinsic::x86_avx512_mask_gather3siv4_si;
10729       break;
10730     case X86::BI__builtin_ia32_gather3siv8sf:
10731       IID = Intrinsic::x86_avx512_mask_gather3siv8_sf;
10732       break;
10733     case X86::BI__builtin_ia32_gather3siv8si:
10734       IID = Intrinsic::x86_avx512_mask_gather3siv8_si;
10735       break;
10736     case X86::BI__builtin_ia32_gathersiv8df:
10737       IID = Intrinsic::x86_avx512_mask_gather_dpd_512;
10738       break;
10739     case X86::BI__builtin_ia32_gathersiv16sf:
10740       IID = Intrinsic::x86_avx512_mask_gather_dps_512;
10741       break;
10742     case X86::BI__builtin_ia32_gatherdiv8df:
10743       IID = Intrinsic::x86_avx512_mask_gather_qpd_512;
10744       break;
10745     case X86::BI__builtin_ia32_gatherdiv16sf:
10746       IID = Intrinsic::x86_avx512_mask_gather_qps_512;
10747       break;
10748     case X86::BI__builtin_ia32_gathersiv8di:
10749       IID = Intrinsic::x86_avx512_mask_gather_dpq_512;
10750       break;
10751     case X86::BI__builtin_ia32_gathersiv16si:
10752       IID = Intrinsic::x86_avx512_mask_gather_dpi_512;
10753       break;
10754     case X86::BI__builtin_ia32_gatherdiv8di:
10755       IID = Intrinsic::x86_avx512_mask_gather_qpq_512;
10756       break;
10757     case X86::BI__builtin_ia32_gatherdiv16si:
10758       IID = Intrinsic::x86_avx512_mask_gather_qpi_512;
10759       break;
10760     }
10761 
10762     unsigned MinElts = std::min(Ops[0]->getType()->getVectorNumElements(),
10763                                 Ops[2]->getType()->getVectorNumElements());
10764     Ops[3] = getMaskVecValue(*this, Ops[3], MinElts);
10765     Function *Intr = CGM.getIntrinsic(IID);
10766     return Builder.CreateCall(Intr, Ops);
10767   }
10768 
10769   case X86::BI__builtin_ia32_scattersiv8df:
10770   case X86::BI__builtin_ia32_scattersiv16sf:
10771   case X86::BI__builtin_ia32_scatterdiv8df:
10772   case X86::BI__builtin_ia32_scatterdiv16sf:
10773   case X86::BI__builtin_ia32_scattersiv8di:
10774   case X86::BI__builtin_ia32_scattersiv16si:
10775   case X86::BI__builtin_ia32_scatterdiv8di:
10776   case X86::BI__builtin_ia32_scatterdiv16si:
10777   case X86::BI__builtin_ia32_scatterdiv2df:
10778   case X86::BI__builtin_ia32_scatterdiv2di:
10779   case X86::BI__builtin_ia32_scatterdiv4df:
10780   case X86::BI__builtin_ia32_scatterdiv4di:
10781   case X86::BI__builtin_ia32_scatterdiv4sf:
10782   case X86::BI__builtin_ia32_scatterdiv4si:
10783   case X86::BI__builtin_ia32_scatterdiv8sf:
10784   case X86::BI__builtin_ia32_scatterdiv8si:
10785   case X86::BI__builtin_ia32_scattersiv2df:
10786   case X86::BI__builtin_ia32_scattersiv2di:
10787   case X86::BI__builtin_ia32_scattersiv4df:
10788   case X86::BI__builtin_ia32_scattersiv4di:
10789   case X86::BI__builtin_ia32_scattersiv4sf:
10790   case X86::BI__builtin_ia32_scattersiv4si:
10791   case X86::BI__builtin_ia32_scattersiv8sf:
10792   case X86::BI__builtin_ia32_scattersiv8si: {
10793     Intrinsic::ID IID;
10794     switch (BuiltinID) {
10795     default: llvm_unreachable("Unexpected builtin");
10796     case X86::BI__builtin_ia32_scattersiv8df:
10797       IID = Intrinsic::x86_avx512_mask_scatter_dpd_512;
10798       break;
10799     case X86::BI__builtin_ia32_scattersiv16sf:
10800       IID = Intrinsic::x86_avx512_mask_scatter_dps_512;
10801       break;
10802     case X86::BI__builtin_ia32_scatterdiv8df:
10803       IID = Intrinsic::x86_avx512_mask_scatter_qpd_512;
10804       break;
10805     case X86::BI__builtin_ia32_scatterdiv16sf:
10806       IID = Intrinsic::x86_avx512_mask_scatter_qps_512;
10807       break;
10808     case X86::BI__builtin_ia32_scattersiv8di:
10809       IID = Intrinsic::x86_avx512_mask_scatter_dpq_512;
10810       break;
10811     case X86::BI__builtin_ia32_scattersiv16si:
10812       IID = Intrinsic::x86_avx512_mask_scatter_dpi_512;
10813       break;
10814     case X86::BI__builtin_ia32_scatterdiv8di:
10815       IID = Intrinsic::x86_avx512_mask_scatter_qpq_512;
10816       break;
10817     case X86::BI__builtin_ia32_scatterdiv16si:
10818       IID = Intrinsic::x86_avx512_mask_scatter_qpi_512;
10819       break;
10820     case X86::BI__builtin_ia32_scatterdiv2df:
10821       IID = Intrinsic::x86_avx512_mask_scatterdiv2_df;
10822       break;
10823     case X86::BI__builtin_ia32_scatterdiv2di:
10824       IID = Intrinsic::x86_avx512_mask_scatterdiv2_di;
10825       break;
10826     case X86::BI__builtin_ia32_scatterdiv4df:
10827       IID = Intrinsic::x86_avx512_mask_scatterdiv4_df;
10828       break;
10829     case X86::BI__builtin_ia32_scatterdiv4di:
10830       IID = Intrinsic::x86_avx512_mask_scatterdiv4_di;
10831       break;
10832     case X86::BI__builtin_ia32_scatterdiv4sf:
10833       IID = Intrinsic::x86_avx512_mask_scatterdiv4_sf;
10834       break;
10835     case X86::BI__builtin_ia32_scatterdiv4si:
10836       IID = Intrinsic::x86_avx512_mask_scatterdiv4_si;
10837       break;
10838     case X86::BI__builtin_ia32_scatterdiv8sf:
10839       IID = Intrinsic::x86_avx512_mask_scatterdiv8_sf;
10840       break;
10841     case X86::BI__builtin_ia32_scatterdiv8si:
10842       IID = Intrinsic::x86_avx512_mask_scatterdiv8_si;
10843       break;
10844     case X86::BI__builtin_ia32_scattersiv2df:
10845       IID = Intrinsic::x86_avx512_mask_scattersiv2_df;
10846       break;
10847     case X86::BI__builtin_ia32_scattersiv2di:
10848       IID = Intrinsic::x86_avx512_mask_scattersiv2_di;
10849       break;
10850     case X86::BI__builtin_ia32_scattersiv4df:
10851       IID = Intrinsic::x86_avx512_mask_scattersiv4_df;
10852       break;
10853     case X86::BI__builtin_ia32_scattersiv4di:
10854       IID = Intrinsic::x86_avx512_mask_scattersiv4_di;
10855       break;
10856     case X86::BI__builtin_ia32_scattersiv4sf:
10857       IID = Intrinsic::x86_avx512_mask_scattersiv4_sf;
10858       break;
10859     case X86::BI__builtin_ia32_scattersiv4si:
10860       IID = Intrinsic::x86_avx512_mask_scattersiv4_si;
10861       break;
10862     case X86::BI__builtin_ia32_scattersiv8sf:
10863       IID = Intrinsic::x86_avx512_mask_scattersiv8_sf;
10864       break;
10865     case X86::BI__builtin_ia32_scattersiv8si:
10866       IID = Intrinsic::x86_avx512_mask_scattersiv8_si;
10867       break;
10868     }
10869 
10870     unsigned MinElts = std::min(Ops[2]->getType()->getVectorNumElements(),
10871                                 Ops[3]->getType()->getVectorNumElements());
10872     Ops[1] = getMaskVecValue(*this, Ops[1], MinElts);
10873     Function *Intr = CGM.getIntrinsic(IID);
10874     return Builder.CreateCall(Intr, Ops);
10875   }
10876 
10877   case X86::BI__builtin_ia32_vextractf128_pd256:
10878   case X86::BI__builtin_ia32_vextractf128_ps256:
10879   case X86::BI__builtin_ia32_vextractf128_si256:
10880   case X86::BI__builtin_ia32_extract128i256:
10881   case X86::BI__builtin_ia32_extractf64x4_mask:
10882   case X86::BI__builtin_ia32_extractf32x4_mask:
10883   case X86::BI__builtin_ia32_extracti64x4_mask:
10884   case X86::BI__builtin_ia32_extracti32x4_mask:
10885   case X86::BI__builtin_ia32_extractf32x8_mask:
10886   case X86::BI__builtin_ia32_extracti32x8_mask:
10887   case X86::BI__builtin_ia32_extractf32x4_256_mask:
10888   case X86::BI__builtin_ia32_extracti32x4_256_mask:
10889   case X86::BI__builtin_ia32_extractf64x2_256_mask:
10890   case X86::BI__builtin_ia32_extracti64x2_256_mask:
10891   case X86::BI__builtin_ia32_extractf64x2_512_mask:
10892   case X86::BI__builtin_ia32_extracti64x2_512_mask: {
10893     llvm::Type *DstTy = ConvertType(E->getType());
10894     unsigned NumElts = DstTy->getVectorNumElements();
10895     unsigned SrcNumElts = Ops[0]->getType()->getVectorNumElements();
10896     unsigned SubVectors = SrcNumElts / NumElts;
10897     unsigned Index = cast<ConstantInt>(Ops[1])->getZExtValue();
10898     assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors");
10899     Index &= SubVectors - 1; // Remove any extra bits.
10900     Index *= NumElts;
10901 
10902     uint32_t Indices[16];
10903     for (unsigned i = 0; i != NumElts; ++i)
10904       Indices[i] = i + Index;
10905 
10906     Value *Res = Builder.CreateShuffleVector(Ops[0],
10907                                              UndefValue::get(Ops[0]->getType()),
10908                                              makeArrayRef(Indices, NumElts),
10909                                              "extract");
10910 
10911     if (Ops.size() == 4)
10912       Res = EmitX86Select(*this, Ops[3], Res, Ops[2]);
10913 
10914     return Res;
10915   }
10916   case X86::BI__builtin_ia32_vinsertf128_pd256:
10917   case X86::BI__builtin_ia32_vinsertf128_ps256:
10918   case X86::BI__builtin_ia32_vinsertf128_si256:
10919   case X86::BI__builtin_ia32_insert128i256:
10920   case X86::BI__builtin_ia32_insertf64x4:
10921   case X86::BI__builtin_ia32_insertf32x4:
10922   case X86::BI__builtin_ia32_inserti64x4:
10923   case X86::BI__builtin_ia32_inserti32x4:
10924   case X86::BI__builtin_ia32_insertf32x8:
10925   case X86::BI__builtin_ia32_inserti32x8:
10926   case X86::BI__builtin_ia32_insertf32x4_256:
10927   case X86::BI__builtin_ia32_inserti32x4_256:
10928   case X86::BI__builtin_ia32_insertf64x2_256:
10929   case X86::BI__builtin_ia32_inserti64x2_256:
10930   case X86::BI__builtin_ia32_insertf64x2_512:
10931   case X86::BI__builtin_ia32_inserti64x2_512: {
10932     unsigned DstNumElts = Ops[0]->getType()->getVectorNumElements();
10933     unsigned SrcNumElts = Ops[1]->getType()->getVectorNumElements();
10934     unsigned SubVectors = DstNumElts / SrcNumElts;
10935     unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue();
10936     assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors");
10937     Index &= SubVectors - 1; // Remove any extra bits.
10938     Index *= SrcNumElts;
10939 
10940     uint32_t Indices[16];
10941     for (unsigned i = 0; i != DstNumElts; ++i)
10942       Indices[i] = (i >= SrcNumElts) ? SrcNumElts + (i % SrcNumElts) : i;
10943 
10944     Value *Op1 = Builder.CreateShuffleVector(Ops[1],
10945                                              UndefValue::get(Ops[1]->getType()),
10946                                              makeArrayRef(Indices, DstNumElts),
10947                                              "widen");
10948 
10949     for (unsigned i = 0; i != DstNumElts; ++i) {
10950       if (i >= Index && i < (Index + SrcNumElts))
10951         Indices[i] = (i - Index) + DstNumElts;
10952       else
10953         Indices[i] = i;
10954     }
10955 
10956     return Builder.CreateShuffleVector(Ops[0], Op1,
10957                                        makeArrayRef(Indices, DstNumElts),
10958                                        "insert");
10959   }
10960   case X86::BI__builtin_ia32_pmovqd512_mask:
10961   case X86::BI__builtin_ia32_pmovwb512_mask: {
10962     Value *Res = Builder.CreateTrunc(Ops[0], Ops[1]->getType());
10963     return EmitX86Select(*this, Ops[2], Res, Ops[1]);
10964   }
10965   case X86::BI__builtin_ia32_pmovdb512_mask:
10966   case X86::BI__builtin_ia32_pmovdw512_mask:
10967   case X86::BI__builtin_ia32_pmovqw512_mask: {
10968     if (const auto *C = dyn_cast<Constant>(Ops[2]))
10969       if (C->isAllOnesValue())
10970         return Builder.CreateTrunc(Ops[0], Ops[1]->getType());
10971 
10972     Intrinsic::ID IID;
10973     switch (BuiltinID) {
10974     default: llvm_unreachable("Unsupported intrinsic!");
10975     case X86::BI__builtin_ia32_pmovdb512_mask:
10976       IID = Intrinsic::x86_avx512_mask_pmov_db_512;
10977       break;
10978     case X86::BI__builtin_ia32_pmovdw512_mask:
10979       IID = Intrinsic::x86_avx512_mask_pmov_dw_512;
10980       break;
10981     case X86::BI__builtin_ia32_pmovqw512_mask:
10982       IID = Intrinsic::x86_avx512_mask_pmov_qw_512;
10983       break;
10984     }
10985 
10986     Function *Intr = CGM.getIntrinsic(IID);
10987     return Builder.CreateCall(Intr, Ops);
10988   }
10989   case X86::BI__builtin_ia32_pblendw128:
10990   case X86::BI__builtin_ia32_blendpd:
10991   case X86::BI__builtin_ia32_blendps:
10992   case X86::BI__builtin_ia32_blendpd256:
10993   case X86::BI__builtin_ia32_blendps256:
10994   case X86::BI__builtin_ia32_pblendw256:
10995   case X86::BI__builtin_ia32_pblendd128:
10996   case X86::BI__builtin_ia32_pblendd256: {
10997     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
10998     unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
10999 
11000     uint32_t Indices[16];
11001     // If there are more than 8 elements, the immediate is used twice so make
11002     // sure we handle that.
11003     for (unsigned i = 0; i != NumElts; ++i)
11004       Indices[i] = ((Imm >> (i % 8)) & 0x1) ? NumElts + i : i;
11005 
11006     return Builder.CreateShuffleVector(Ops[0], Ops[1],
11007                                        makeArrayRef(Indices, NumElts),
11008                                        "blend");
11009   }
11010   case X86::BI__builtin_ia32_pshuflw:
11011   case X86::BI__builtin_ia32_pshuflw256:
11012   case X86::BI__builtin_ia32_pshuflw512: {
11013     uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
11014     llvm::Type *Ty = Ops[0]->getType();
11015     unsigned NumElts = Ty->getVectorNumElements();
11016 
11017     // Splat the 8-bits of immediate 4 times to help the loop wrap around.
11018     Imm = (Imm & 0xff) * 0x01010101;
11019 
11020     uint32_t Indices[32];
11021     for (unsigned l = 0; l != NumElts; l += 8) {
11022       for (unsigned i = 0; i != 4; ++i) {
11023         Indices[l + i] = l + (Imm & 3);
11024         Imm >>= 2;
11025       }
11026       for (unsigned i = 4; i != 8; ++i)
11027         Indices[l + i] = l + i;
11028     }
11029 
11030     return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty),
11031                                        makeArrayRef(Indices, NumElts),
11032                                        "pshuflw");
11033   }
11034   case X86::BI__builtin_ia32_pshufhw:
11035   case X86::BI__builtin_ia32_pshufhw256:
11036   case X86::BI__builtin_ia32_pshufhw512: {
11037     uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
11038     llvm::Type *Ty = Ops[0]->getType();
11039     unsigned NumElts = Ty->getVectorNumElements();
11040 
11041     // Splat the 8-bits of immediate 4 times to help the loop wrap around.
11042     Imm = (Imm & 0xff) * 0x01010101;
11043 
11044     uint32_t Indices[32];
11045     for (unsigned l = 0; l != NumElts; l += 8) {
11046       for (unsigned i = 0; i != 4; ++i)
11047         Indices[l + i] = l + i;
11048       for (unsigned i = 4; i != 8; ++i) {
11049         Indices[l + i] = l + 4 + (Imm & 3);
11050         Imm >>= 2;
11051       }
11052     }
11053 
11054     return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty),
11055                                        makeArrayRef(Indices, NumElts),
11056                                        "pshufhw");
11057   }
11058   case X86::BI__builtin_ia32_pshufd:
11059   case X86::BI__builtin_ia32_pshufd256:
11060   case X86::BI__builtin_ia32_pshufd512:
11061   case X86::BI__builtin_ia32_vpermilpd:
11062   case X86::BI__builtin_ia32_vpermilps:
11063   case X86::BI__builtin_ia32_vpermilpd256:
11064   case X86::BI__builtin_ia32_vpermilps256:
11065   case X86::BI__builtin_ia32_vpermilpd512:
11066   case X86::BI__builtin_ia32_vpermilps512: {
11067     uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
11068     llvm::Type *Ty = Ops[0]->getType();
11069     unsigned NumElts = Ty->getVectorNumElements();
11070     unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128;
11071     unsigned NumLaneElts = NumElts / NumLanes;
11072 
11073     // Splat the 8-bits of immediate 4 times to help the loop wrap around.
11074     Imm = (Imm & 0xff) * 0x01010101;
11075 
11076     uint32_t Indices[16];
11077     for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
11078       for (unsigned i = 0; i != NumLaneElts; ++i) {
11079         Indices[i + l] = (Imm % NumLaneElts) + l;
11080         Imm /= NumLaneElts;
11081       }
11082     }
11083 
11084     return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty),
11085                                        makeArrayRef(Indices, NumElts),
11086                                        "permil");
11087   }
11088   case X86::BI__builtin_ia32_shufpd:
11089   case X86::BI__builtin_ia32_shufpd256:
11090   case X86::BI__builtin_ia32_shufpd512:
11091   case X86::BI__builtin_ia32_shufps:
11092   case X86::BI__builtin_ia32_shufps256:
11093   case X86::BI__builtin_ia32_shufps512: {
11094     uint32_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
11095     llvm::Type *Ty = Ops[0]->getType();
11096     unsigned NumElts = Ty->getVectorNumElements();
11097     unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128;
11098     unsigned NumLaneElts = NumElts / NumLanes;
11099 
11100     // Splat the 8-bits of immediate 4 times to help the loop wrap around.
11101     Imm = (Imm & 0xff) * 0x01010101;
11102 
11103     uint32_t Indices[16];
11104     for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
11105       for (unsigned i = 0; i != NumLaneElts; ++i) {
11106         unsigned Index = Imm % NumLaneElts;
11107         Imm /= NumLaneElts;
11108         if (i >= (NumLaneElts / 2))
11109           Index += NumElts;
11110         Indices[l + i] = l + Index;
11111       }
11112     }
11113 
11114     return Builder.CreateShuffleVector(Ops[0], Ops[1],
11115                                        makeArrayRef(Indices, NumElts),
11116                                        "shufp");
11117   }
11118   case X86::BI__builtin_ia32_permdi256:
11119   case X86::BI__builtin_ia32_permdf256:
11120   case X86::BI__builtin_ia32_permdi512:
11121   case X86::BI__builtin_ia32_permdf512: {
11122     unsigned Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
11123     llvm::Type *Ty = Ops[0]->getType();
11124     unsigned NumElts = Ty->getVectorNumElements();
11125 
11126     // These intrinsics operate on 256-bit lanes of four 64-bit elements.
11127     uint32_t Indices[8];
11128     for (unsigned l = 0; l != NumElts; l += 4)
11129       for (unsigned i = 0; i != 4; ++i)
11130         Indices[l + i] = l + ((Imm >> (2 * i)) & 0x3);
11131 
11132     return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty),
11133                                        makeArrayRef(Indices, NumElts),
11134                                        "perm");
11135   }
11136   case X86::BI__builtin_ia32_palignr128:
11137   case X86::BI__builtin_ia32_palignr256:
11138   case X86::BI__builtin_ia32_palignr512: {
11139     unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff;
11140 
11141     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
11142     assert(NumElts % 16 == 0);
11143 
11144     // If palignr is shifting the pair of vectors more than the size of two
11145     // lanes, emit zero.
11146     if (ShiftVal >= 32)
11147       return llvm::Constant::getNullValue(ConvertType(E->getType()));
11148 
11149     // If palignr is shifting the pair of input vectors more than one lane,
11150     // but less than two lanes, convert to shifting in zeroes.
11151     if (ShiftVal > 16) {
11152       ShiftVal -= 16;
11153       Ops[1] = Ops[0];
11154       Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType());
11155     }
11156 
11157     uint32_t Indices[64];
11158     // 256-bit palignr operates on 128-bit lanes so we need to handle that
11159     for (unsigned l = 0; l != NumElts; l += 16) {
11160       for (unsigned i = 0; i != 16; ++i) {
11161         unsigned Idx = ShiftVal + i;
11162         if (Idx >= 16)
11163           Idx += NumElts - 16; // End of lane, switch operand.
11164         Indices[l + i] = Idx + l;
11165       }
11166     }
11167 
11168     return Builder.CreateShuffleVector(Ops[1], Ops[0],
11169                                        makeArrayRef(Indices, NumElts),
11170                                        "palignr");
11171   }
11172   case X86::BI__builtin_ia32_alignd128:
11173   case X86::BI__builtin_ia32_alignd256:
11174   case X86::BI__builtin_ia32_alignd512:
11175   case X86::BI__builtin_ia32_alignq128:
11176   case X86::BI__builtin_ia32_alignq256:
11177   case X86::BI__builtin_ia32_alignq512: {
11178     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
11179     unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff;
11180 
11181     // Mask the shift amount to width of two vectors.
11182     ShiftVal &= (2 * NumElts) - 1;
11183 
11184     uint32_t Indices[16];
11185     for (unsigned i = 0; i != NumElts; ++i)
11186       Indices[i] = i + ShiftVal;
11187 
11188     return Builder.CreateShuffleVector(Ops[1], Ops[0],
11189                                        makeArrayRef(Indices, NumElts),
11190                                        "valign");
11191   }
11192   case X86::BI__builtin_ia32_shuf_f32x4_256:
11193   case X86::BI__builtin_ia32_shuf_f64x2_256:
11194   case X86::BI__builtin_ia32_shuf_i32x4_256:
11195   case X86::BI__builtin_ia32_shuf_i64x2_256:
11196   case X86::BI__builtin_ia32_shuf_f32x4:
11197   case X86::BI__builtin_ia32_shuf_f64x2:
11198   case X86::BI__builtin_ia32_shuf_i32x4:
11199   case X86::BI__builtin_ia32_shuf_i64x2: {
11200     unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
11201     llvm::Type *Ty = Ops[0]->getType();
11202     unsigned NumElts = Ty->getVectorNumElements();
11203     unsigned NumLanes = Ty->getPrimitiveSizeInBits() == 512 ? 4 : 2;
11204     unsigned NumLaneElts = NumElts / NumLanes;
11205 
11206     uint32_t Indices[16];
11207     for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
11208       unsigned Index = (Imm % NumLanes) * NumLaneElts;
11209       Imm /= NumLanes; // Discard the bits we just used.
11210       if (l >= (NumElts / 2))
11211         Index += NumElts; // Switch to other source.
11212       for (unsigned i = 0; i != NumLaneElts; ++i) {
11213         Indices[l + i] = Index + i;
11214       }
11215     }
11216 
11217     return Builder.CreateShuffleVector(Ops[0], Ops[1],
11218                                        makeArrayRef(Indices, NumElts),
11219                                        "shuf");
11220   }
11221 
11222   case X86::BI__builtin_ia32_vperm2f128_pd256:
11223   case X86::BI__builtin_ia32_vperm2f128_ps256:
11224   case X86::BI__builtin_ia32_vperm2f128_si256:
11225   case X86::BI__builtin_ia32_permti256: {
11226     unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
11227     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
11228 
11229     // This takes a very simple approach since there are two lanes and a
11230     // shuffle can have 2 inputs. So we reserve the first input for the first
11231     // lane and the second input for the second lane. This may result in
11232     // duplicate sources, but this can be dealt with in the backend.
11233 
11234     Value *OutOps[2];
11235     uint32_t Indices[8];
11236     for (unsigned l = 0; l != 2; ++l) {
11237       // Determine the source for this lane.
11238       if (Imm & (1 << ((l * 4) + 3)))
11239         OutOps[l] = llvm::ConstantAggregateZero::get(Ops[0]->getType());
11240       else if (Imm & (1 << ((l * 4) + 1)))
11241         OutOps[l] = Ops[1];
11242       else
11243         OutOps[l] = Ops[0];
11244 
11245       for (unsigned i = 0; i != NumElts/2; ++i) {
11246         // Start with ith element of the source for this lane.
11247         unsigned Idx = (l * NumElts) + i;
11248         // If bit 0 of the immediate half is set, switch to the high half of
11249         // the source.
11250         if (Imm & (1 << (l * 4)))
11251           Idx += NumElts/2;
11252         Indices[(l * (NumElts/2)) + i] = Idx;
11253       }
11254     }
11255 
11256     return Builder.CreateShuffleVector(OutOps[0], OutOps[1],
11257                                        makeArrayRef(Indices, NumElts),
11258                                        "vperm");
11259   }
11260 
11261   case X86::BI__builtin_ia32_pslldqi128_byteshift:
11262   case X86::BI__builtin_ia32_pslldqi256_byteshift:
11263   case X86::BI__builtin_ia32_pslldqi512_byteshift: {
11264     unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
11265     llvm::Type *ResultType = Ops[0]->getType();
11266     // Builtin type is vXi64 so multiply by 8 to get bytes.
11267     unsigned NumElts = ResultType->getVectorNumElements() * 8;
11268 
11269     // If pslldq is shifting the vector more than 15 bytes, emit zero.
11270     if (ShiftVal >= 16)
11271       return llvm::Constant::getNullValue(ResultType);
11272 
11273     uint32_t Indices[64];
11274     // 256/512-bit pslldq operates on 128-bit lanes so we need to handle that
11275     for (unsigned l = 0; l != NumElts; l += 16) {
11276       for (unsigned i = 0; i != 16; ++i) {
11277         unsigned Idx = NumElts + i - ShiftVal;
11278         if (Idx < NumElts) Idx -= NumElts - 16; // end of lane, switch operand.
11279         Indices[l + i] = Idx + l;
11280       }
11281     }
11282 
11283     llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, NumElts);
11284     Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast");
11285     Value *Zero = llvm::Constant::getNullValue(VecTy);
11286     Value *SV = Builder.CreateShuffleVector(Zero, Cast,
11287                                             makeArrayRef(Indices, NumElts),
11288                                             "pslldq");
11289     return Builder.CreateBitCast(SV, Ops[0]->getType(), "cast");
11290   }
11291   case X86::BI__builtin_ia32_psrldqi128_byteshift:
11292   case X86::BI__builtin_ia32_psrldqi256_byteshift:
11293   case X86::BI__builtin_ia32_psrldqi512_byteshift: {
11294     unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
11295     llvm::Type *ResultType = Ops[0]->getType();
11296     // Builtin type is vXi64 so multiply by 8 to get bytes.
11297     unsigned NumElts = ResultType->getVectorNumElements() * 8;
11298 
11299     // If psrldq is shifting the vector more than 15 bytes, emit zero.
11300     if (ShiftVal >= 16)
11301       return llvm::Constant::getNullValue(ResultType);
11302 
11303     uint32_t Indices[64];
11304     // 256/512-bit psrldq operates on 128-bit lanes so we need to handle that
11305     for (unsigned l = 0; l != NumElts; l += 16) {
11306       for (unsigned i = 0; i != 16; ++i) {
11307         unsigned Idx = i + ShiftVal;
11308         if (Idx >= 16) Idx += NumElts - 16; // end of lane, switch operand.
11309         Indices[l + i] = Idx + l;
11310       }
11311     }
11312 
11313     llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, NumElts);
11314     Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast");
11315     Value *Zero = llvm::Constant::getNullValue(VecTy);
11316     Value *SV = Builder.CreateShuffleVector(Cast, Zero,
11317                                             makeArrayRef(Indices, NumElts),
11318                                             "psrldq");
11319     return Builder.CreateBitCast(SV, ResultType, "cast");
11320   }
11321   case X86::BI__builtin_ia32_kshiftliqi:
11322   case X86::BI__builtin_ia32_kshiftlihi:
11323   case X86::BI__builtin_ia32_kshiftlisi:
11324   case X86::BI__builtin_ia32_kshiftlidi: {
11325     unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
11326     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11327 
11328     if (ShiftVal >= NumElts)
11329       return llvm::Constant::getNullValue(Ops[0]->getType());
11330 
11331     Value *In = getMaskVecValue(*this, Ops[0], NumElts);
11332 
11333     uint32_t Indices[64];
11334     for (unsigned i = 0; i != NumElts; ++i)
11335       Indices[i] = NumElts + i - ShiftVal;
11336 
11337     Value *Zero = llvm::Constant::getNullValue(In->getType());
11338     Value *SV = Builder.CreateShuffleVector(Zero, In,
11339                                             makeArrayRef(Indices, NumElts),
11340                                             "kshiftl");
11341     return Builder.CreateBitCast(SV, Ops[0]->getType());
11342   }
11343   case X86::BI__builtin_ia32_kshiftriqi:
11344   case X86::BI__builtin_ia32_kshiftrihi:
11345   case X86::BI__builtin_ia32_kshiftrisi:
11346   case X86::BI__builtin_ia32_kshiftridi: {
11347     unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
11348     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11349 
11350     if (ShiftVal >= NumElts)
11351       return llvm::Constant::getNullValue(Ops[0]->getType());
11352 
11353     Value *In = getMaskVecValue(*this, Ops[0], NumElts);
11354 
11355     uint32_t Indices[64];
11356     for (unsigned i = 0; i != NumElts; ++i)
11357       Indices[i] = i + ShiftVal;
11358 
11359     Value *Zero = llvm::Constant::getNullValue(In->getType());
11360     Value *SV = Builder.CreateShuffleVector(In, Zero,
11361                                             makeArrayRef(Indices, NumElts),
11362                                             "kshiftr");
11363     return Builder.CreateBitCast(SV, Ops[0]->getType());
11364   }
11365   case X86::BI__builtin_ia32_movnti:
11366   case X86::BI__builtin_ia32_movnti64:
11367   case X86::BI__builtin_ia32_movntsd:
11368   case X86::BI__builtin_ia32_movntss: {
11369     llvm::MDNode *Node = llvm::MDNode::get(
11370         getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1)));
11371 
11372     Value *Ptr = Ops[0];
11373     Value *Src = Ops[1];
11374 
11375     // Extract the 0'th element of the source vector.
11376     if (BuiltinID == X86::BI__builtin_ia32_movntsd ||
11377         BuiltinID == X86::BI__builtin_ia32_movntss)
11378       Src = Builder.CreateExtractElement(Src, (uint64_t)0, "extract");
11379 
11380     // Convert the type of the pointer to a pointer to the stored type.
11381     Value *BC = Builder.CreateBitCast(
11382         Ptr, llvm::PointerType::getUnqual(Src->getType()), "cast");
11383 
11384     // Unaligned nontemporal store of the scalar value.
11385     StoreInst *SI = Builder.CreateDefaultAlignedStore(Src, BC);
11386     SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
11387     SI->setAlignment(llvm::Align::None());
11388     return SI;
11389   }
11390   // Rotate is a special case of funnel shift - 1st 2 args are the same.
11391   case X86::BI__builtin_ia32_vprotb:
11392   case X86::BI__builtin_ia32_vprotw:
11393   case X86::BI__builtin_ia32_vprotd:
11394   case X86::BI__builtin_ia32_vprotq:
11395   case X86::BI__builtin_ia32_vprotbi:
11396   case X86::BI__builtin_ia32_vprotwi:
11397   case X86::BI__builtin_ia32_vprotdi:
11398   case X86::BI__builtin_ia32_vprotqi:
11399   case X86::BI__builtin_ia32_prold128:
11400   case X86::BI__builtin_ia32_prold256:
11401   case X86::BI__builtin_ia32_prold512:
11402   case X86::BI__builtin_ia32_prolq128:
11403   case X86::BI__builtin_ia32_prolq256:
11404   case X86::BI__builtin_ia32_prolq512:
11405   case X86::BI__builtin_ia32_prolvd128:
11406   case X86::BI__builtin_ia32_prolvd256:
11407   case X86::BI__builtin_ia32_prolvd512:
11408   case X86::BI__builtin_ia32_prolvq128:
11409   case X86::BI__builtin_ia32_prolvq256:
11410   case X86::BI__builtin_ia32_prolvq512:
11411     return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], false);
11412   case X86::BI__builtin_ia32_prord128:
11413   case X86::BI__builtin_ia32_prord256:
11414   case X86::BI__builtin_ia32_prord512:
11415   case X86::BI__builtin_ia32_prorq128:
11416   case X86::BI__builtin_ia32_prorq256:
11417   case X86::BI__builtin_ia32_prorq512:
11418   case X86::BI__builtin_ia32_prorvd128:
11419   case X86::BI__builtin_ia32_prorvd256:
11420   case X86::BI__builtin_ia32_prorvd512:
11421   case X86::BI__builtin_ia32_prorvq128:
11422   case X86::BI__builtin_ia32_prorvq256:
11423   case X86::BI__builtin_ia32_prorvq512:
11424     return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], true);
11425   case X86::BI__builtin_ia32_selectb_128:
11426   case X86::BI__builtin_ia32_selectb_256:
11427   case X86::BI__builtin_ia32_selectb_512:
11428   case X86::BI__builtin_ia32_selectw_128:
11429   case X86::BI__builtin_ia32_selectw_256:
11430   case X86::BI__builtin_ia32_selectw_512:
11431   case X86::BI__builtin_ia32_selectd_128:
11432   case X86::BI__builtin_ia32_selectd_256:
11433   case X86::BI__builtin_ia32_selectd_512:
11434   case X86::BI__builtin_ia32_selectq_128:
11435   case X86::BI__builtin_ia32_selectq_256:
11436   case X86::BI__builtin_ia32_selectq_512:
11437   case X86::BI__builtin_ia32_selectps_128:
11438   case X86::BI__builtin_ia32_selectps_256:
11439   case X86::BI__builtin_ia32_selectps_512:
11440   case X86::BI__builtin_ia32_selectpd_128:
11441   case X86::BI__builtin_ia32_selectpd_256:
11442   case X86::BI__builtin_ia32_selectpd_512:
11443     return EmitX86Select(*this, Ops[0], Ops[1], Ops[2]);
11444   case X86::BI__builtin_ia32_selectss_128:
11445   case X86::BI__builtin_ia32_selectsd_128: {
11446     Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
11447     Value *B = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
11448     A = EmitX86ScalarSelect(*this, Ops[0], A, B);
11449     return Builder.CreateInsertElement(Ops[1], A, (uint64_t)0);
11450   }
11451   case X86::BI__builtin_ia32_cmpb128_mask:
11452   case X86::BI__builtin_ia32_cmpb256_mask:
11453   case X86::BI__builtin_ia32_cmpb512_mask:
11454   case X86::BI__builtin_ia32_cmpw128_mask:
11455   case X86::BI__builtin_ia32_cmpw256_mask:
11456   case X86::BI__builtin_ia32_cmpw512_mask:
11457   case X86::BI__builtin_ia32_cmpd128_mask:
11458   case X86::BI__builtin_ia32_cmpd256_mask:
11459   case X86::BI__builtin_ia32_cmpd512_mask:
11460   case X86::BI__builtin_ia32_cmpq128_mask:
11461   case X86::BI__builtin_ia32_cmpq256_mask:
11462   case X86::BI__builtin_ia32_cmpq512_mask: {
11463     unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
11464     return EmitX86MaskedCompare(*this, CC, true, Ops);
11465   }
11466   case X86::BI__builtin_ia32_ucmpb128_mask:
11467   case X86::BI__builtin_ia32_ucmpb256_mask:
11468   case X86::BI__builtin_ia32_ucmpb512_mask:
11469   case X86::BI__builtin_ia32_ucmpw128_mask:
11470   case X86::BI__builtin_ia32_ucmpw256_mask:
11471   case X86::BI__builtin_ia32_ucmpw512_mask:
11472   case X86::BI__builtin_ia32_ucmpd128_mask:
11473   case X86::BI__builtin_ia32_ucmpd256_mask:
11474   case X86::BI__builtin_ia32_ucmpd512_mask:
11475   case X86::BI__builtin_ia32_ucmpq128_mask:
11476   case X86::BI__builtin_ia32_ucmpq256_mask:
11477   case X86::BI__builtin_ia32_ucmpq512_mask: {
11478     unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
11479     return EmitX86MaskedCompare(*this, CC, false, Ops);
11480   }
11481   case X86::BI__builtin_ia32_vpcomb:
11482   case X86::BI__builtin_ia32_vpcomw:
11483   case X86::BI__builtin_ia32_vpcomd:
11484   case X86::BI__builtin_ia32_vpcomq:
11485     return EmitX86vpcom(*this, Ops, true);
11486   case X86::BI__builtin_ia32_vpcomub:
11487   case X86::BI__builtin_ia32_vpcomuw:
11488   case X86::BI__builtin_ia32_vpcomud:
11489   case X86::BI__builtin_ia32_vpcomuq:
11490     return EmitX86vpcom(*this, Ops, false);
11491 
11492   case X86::BI__builtin_ia32_kortestcqi:
11493   case X86::BI__builtin_ia32_kortestchi:
11494   case X86::BI__builtin_ia32_kortestcsi:
11495   case X86::BI__builtin_ia32_kortestcdi: {
11496     Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops);
11497     Value *C = llvm::Constant::getAllOnesValue(Ops[0]->getType());
11498     Value *Cmp = Builder.CreateICmpEQ(Or, C);
11499     return Builder.CreateZExt(Cmp, ConvertType(E->getType()));
11500   }
11501   case X86::BI__builtin_ia32_kortestzqi:
11502   case X86::BI__builtin_ia32_kortestzhi:
11503   case X86::BI__builtin_ia32_kortestzsi:
11504   case X86::BI__builtin_ia32_kortestzdi: {
11505     Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops);
11506     Value *C = llvm::Constant::getNullValue(Ops[0]->getType());
11507     Value *Cmp = Builder.CreateICmpEQ(Or, C);
11508     return Builder.CreateZExt(Cmp, ConvertType(E->getType()));
11509   }
11510 
11511   case X86::BI__builtin_ia32_ktestcqi:
11512   case X86::BI__builtin_ia32_ktestzqi:
11513   case X86::BI__builtin_ia32_ktestchi:
11514   case X86::BI__builtin_ia32_ktestzhi:
11515   case X86::BI__builtin_ia32_ktestcsi:
11516   case X86::BI__builtin_ia32_ktestzsi:
11517   case X86::BI__builtin_ia32_ktestcdi:
11518   case X86::BI__builtin_ia32_ktestzdi: {
11519     Intrinsic::ID IID;
11520     switch (BuiltinID) {
11521     default: llvm_unreachable("Unsupported intrinsic!");
11522     case X86::BI__builtin_ia32_ktestcqi:
11523       IID = Intrinsic::x86_avx512_ktestc_b;
11524       break;
11525     case X86::BI__builtin_ia32_ktestzqi:
11526       IID = Intrinsic::x86_avx512_ktestz_b;
11527       break;
11528     case X86::BI__builtin_ia32_ktestchi:
11529       IID = Intrinsic::x86_avx512_ktestc_w;
11530       break;
11531     case X86::BI__builtin_ia32_ktestzhi:
11532       IID = Intrinsic::x86_avx512_ktestz_w;
11533       break;
11534     case X86::BI__builtin_ia32_ktestcsi:
11535       IID = Intrinsic::x86_avx512_ktestc_d;
11536       break;
11537     case X86::BI__builtin_ia32_ktestzsi:
11538       IID = Intrinsic::x86_avx512_ktestz_d;
11539       break;
11540     case X86::BI__builtin_ia32_ktestcdi:
11541       IID = Intrinsic::x86_avx512_ktestc_q;
11542       break;
11543     case X86::BI__builtin_ia32_ktestzdi:
11544       IID = Intrinsic::x86_avx512_ktestz_q;
11545       break;
11546     }
11547 
11548     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11549     Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
11550     Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
11551     Function *Intr = CGM.getIntrinsic(IID);
11552     return Builder.CreateCall(Intr, {LHS, RHS});
11553   }
11554 
11555   case X86::BI__builtin_ia32_kaddqi:
11556   case X86::BI__builtin_ia32_kaddhi:
11557   case X86::BI__builtin_ia32_kaddsi:
11558   case X86::BI__builtin_ia32_kadddi: {
11559     Intrinsic::ID IID;
11560     switch (BuiltinID) {
11561     default: llvm_unreachable("Unsupported intrinsic!");
11562     case X86::BI__builtin_ia32_kaddqi:
11563       IID = Intrinsic::x86_avx512_kadd_b;
11564       break;
11565     case X86::BI__builtin_ia32_kaddhi:
11566       IID = Intrinsic::x86_avx512_kadd_w;
11567       break;
11568     case X86::BI__builtin_ia32_kaddsi:
11569       IID = Intrinsic::x86_avx512_kadd_d;
11570       break;
11571     case X86::BI__builtin_ia32_kadddi:
11572       IID = Intrinsic::x86_avx512_kadd_q;
11573       break;
11574     }
11575 
11576     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11577     Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
11578     Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
11579     Function *Intr = CGM.getIntrinsic(IID);
11580     Value *Res = Builder.CreateCall(Intr, {LHS, RHS});
11581     return Builder.CreateBitCast(Res, Ops[0]->getType());
11582   }
11583   case X86::BI__builtin_ia32_kandqi:
11584   case X86::BI__builtin_ia32_kandhi:
11585   case X86::BI__builtin_ia32_kandsi:
11586   case X86::BI__builtin_ia32_kanddi:
11587     return EmitX86MaskLogic(*this, Instruction::And, Ops);
11588   case X86::BI__builtin_ia32_kandnqi:
11589   case X86::BI__builtin_ia32_kandnhi:
11590   case X86::BI__builtin_ia32_kandnsi:
11591   case X86::BI__builtin_ia32_kandndi:
11592     return EmitX86MaskLogic(*this, Instruction::And, Ops, true);
11593   case X86::BI__builtin_ia32_korqi:
11594   case X86::BI__builtin_ia32_korhi:
11595   case X86::BI__builtin_ia32_korsi:
11596   case X86::BI__builtin_ia32_kordi:
11597     return EmitX86MaskLogic(*this, Instruction::Or, Ops);
11598   case X86::BI__builtin_ia32_kxnorqi:
11599   case X86::BI__builtin_ia32_kxnorhi:
11600   case X86::BI__builtin_ia32_kxnorsi:
11601   case X86::BI__builtin_ia32_kxnordi:
11602     return EmitX86MaskLogic(*this, Instruction::Xor, Ops, true);
11603   case X86::BI__builtin_ia32_kxorqi:
11604   case X86::BI__builtin_ia32_kxorhi:
11605   case X86::BI__builtin_ia32_kxorsi:
11606   case X86::BI__builtin_ia32_kxordi:
11607     return EmitX86MaskLogic(*this, Instruction::Xor,  Ops);
11608   case X86::BI__builtin_ia32_knotqi:
11609   case X86::BI__builtin_ia32_knothi:
11610   case X86::BI__builtin_ia32_knotsi:
11611   case X86::BI__builtin_ia32_knotdi: {
11612     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11613     Value *Res = getMaskVecValue(*this, Ops[0], NumElts);
11614     return Builder.CreateBitCast(Builder.CreateNot(Res),
11615                                  Ops[0]->getType());
11616   }
11617   case X86::BI__builtin_ia32_kmovb:
11618   case X86::BI__builtin_ia32_kmovw:
11619   case X86::BI__builtin_ia32_kmovd:
11620   case X86::BI__builtin_ia32_kmovq: {
11621     // Bitcast to vXi1 type and then back to integer. This gets the mask
11622     // register type into the IR, but might be optimized out depending on
11623     // what's around it.
11624     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11625     Value *Res = getMaskVecValue(*this, Ops[0], NumElts);
11626     return Builder.CreateBitCast(Res, Ops[0]->getType());
11627   }
11628 
11629   case X86::BI__builtin_ia32_kunpckdi:
11630   case X86::BI__builtin_ia32_kunpcksi:
11631   case X86::BI__builtin_ia32_kunpckhi: {
11632     unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11633     Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
11634     Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
11635     uint32_t Indices[64];
11636     for (unsigned i = 0; i != NumElts; ++i)
11637       Indices[i] = i;
11638 
11639     // First extract half of each vector. This gives better codegen than
11640     // doing it in a single shuffle.
11641     LHS = Builder.CreateShuffleVector(LHS, LHS,
11642                                       makeArrayRef(Indices, NumElts / 2));
11643     RHS = Builder.CreateShuffleVector(RHS, RHS,
11644                                       makeArrayRef(Indices, NumElts / 2));
11645     // Concat the vectors.
11646     // NOTE: Operands are swapped to match the intrinsic definition.
11647     Value *Res = Builder.CreateShuffleVector(RHS, LHS,
11648                                              makeArrayRef(Indices, NumElts));
11649     return Builder.CreateBitCast(Res, Ops[0]->getType());
11650   }
11651 
11652   case X86::BI__builtin_ia32_vplzcntd_128:
11653   case X86::BI__builtin_ia32_vplzcntd_256:
11654   case X86::BI__builtin_ia32_vplzcntd_512:
11655   case X86::BI__builtin_ia32_vplzcntq_128:
11656   case X86::BI__builtin_ia32_vplzcntq_256:
11657   case X86::BI__builtin_ia32_vplzcntq_512: {
11658     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
11659     return Builder.CreateCall(F, {Ops[0],Builder.getInt1(false)});
11660   }
11661   case X86::BI__builtin_ia32_sqrtss:
11662   case X86::BI__builtin_ia32_sqrtsd: {
11663     Value *A = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
11664     Function *F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType());
11665     A = Builder.CreateCall(F, {A});
11666     return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0);
11667   }
11668   case X86::BI__builtin_ia32_sqrtsd_round_mask:
11669   case X86::BI__builtin_ia32_sqrtss_round_mask: {
11670     unsigned CC = cast<llvm::ConstantInt>(Ops[4])->getZExtValue();
11671     // Support only if the rounding mode is 4 (AKA CUR_DIRECTION),
11672     // otherwise keep the intrinsic.
11673     if (CC != 4) {
11674       Intrinsic::ID IID = BuiltinID == X86::BI__builtin_ia32_sqrtsd_round_mask ?
11675                           Intrinsic::x86_avx512_mask_sqrt_sd :
11676                           Intrinsic::x86_avx512_mask_sqrt_ss;
11677       return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
11678     }
11679     Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
11680     Function *F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType());
11681     A = Builder.CreateCall(F, A);
11682     Value *Src = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
11683     A = EmitX86ScalarSelect(*this, Ops[3], A, Src);
11684     return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0);
11685   }
11686   case X86::BI__builtin_ia32_sqrtpd256:
11687   case X86::BI__builtin_ia32_sqrtpd:
11688   case X86::BI__builtin_ia32_sqrtps256:
11689   case X86::BI__builtin_ia32_sqrtps:
11690   case X86::BI__builtin_ia32_sqrtps512:
11691   case X86::BI__builtin_ia32_sqrtpd512: {
11692     if (Ops.size() == 2) {
11693       unsigned CC = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
11694       // Support only if the rounding mode is 4 (AKA CUR_DIRECTION),
11695       // otherwise keep the intrinsic.
11696       if (CC != 4) {
11697         Intrinsic::ID IID = BuiltinID == X86::BI__builtin_ia32_sqrtps512 ?
11698                             Intrinsic::x86_avx512_sqrt_ps_512 :
11699                             Intrinsic::x86_avx512_sqrt_pd_512;
11700         return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
11701       }
11702     }
11703     Function *F = CGM.getIntrinsic(Intrinsic::sqrt, Ops[0]->getType());
11704     return Builder.CreateCall(F, Ops[0]);
11705   }
11706   case X86::BI__builtin_ia32_pabsb128:
11707   case X86::BI__builtin_ia32_pabsw128:
11708   case X86::BI__builtin_ia32_pabsd128:
11709   case X86::BI__builtin_ia32_pabsb256:
11710   case X86::BI__builtin_ia32_pabsw256:
11711   case X86::BI__builtin_ia32_pabsd256:
11712   case X86::BI__builtin_ia32_pabsq128:
11713   case X86::BI__builtin_ia32_pabsq256:
11714   case X86::BI__builtin_ia32_pabsb512:
11715   case X86::BI__builtin_ia32_pabsw512:
11716   case X86::BI__builtin_ia32_pabsd512:
11717   case X86::BI__builtin_ia32_pabsq512:
11718     return EmitX86Abs(*this, Ops);
11719 
11720   case X86::BI__builtin_ia32_pmaxsb128:
11721   case X86::BI__builtin_ia32_pmaxsw128:
11722   case X86::BI__builtin_ia32_pmaxsd128:
11723   case X86::BI__builtin_ia32_pmaxsq128:
11724   case X86::BI__builtin_ia32_pmaxsb256:
11725   case X86::BI__builtin_ia32_pmaxsw256:
11726   case X86::BI__builtin_ia32_pmaxsd256:
11727   case X86::BI__builtin_ia32_pmaxsq256:
11728   case X86::BI__builtin_ia32_pmaxsb512:
11729   case X86::BI__builtin_ia32_pmaxsw512:
11730   case X86::BI__builtin_ia32_pmaxsd512:
11731   case X86::BI__builtin_ia32_pmaxsq512:
11732     return EmitX86MinMax(*this, ICmpInst::ICMP_SGT, Ops);
11733   case X86::BI__builtin_ia32_pmaxub128:
11734   case X86::BI__builtin_ia32_pmaxuw128:
11735   case X86::BI__builtin_ia32_pmaxud128:
11736   case X86::BI__builtin_ia32_pmaxuq128:
11737   case X86::BI__builtin_ia32_pmaxub256:
11738   case X86::BI__builtin_ia32_pmaxuw256:
11739   case X86::BI__builtin_ia32_pmaxud256:
11740   case X86::BI__builtin_ia32_pmaxuq256:
11741   case X86::BI__builtin_ia32_pmaxub512:
11742   case X86::BI__builtin_ia32_pmaxuw512:
11743   case X86::BI__builtin_ia32_pmaxud512:
11744   case X86::BI__builtin_ia32_pmaxuq512:
11745     return EmitX86MinMax(*this, ICmpInst::ICMP_UGT, Ops);
11746   case X86::BI__builtin_ia32_pminsb128:
11747   case X86::BI__builtin_ia32_pminsw128:
11748   case X86::BI__builtin_ia32_pminsd128:
11749   case X86::BI__builtin_ia32_pminsq128:
11750   case X86::BI__builtin_ia32_pminsb256:
11751   case X86::BI__builtin_ia32_pminsw256:
11752   case X86::BI__builtin_ia32_pminsd256:
11753   case X86::BI__builtin_ia32_pminsq256:
11754   case X86::BI__builtin_ia32_pminsb512:
11755   case X86::BI__builtin_ia32_pminsw512:
11756   case X86::BI__builtin_ia32_pminsd512:
11757   case X86::BI__builtin_ia32_pminsq512:
11758     return EmitX86MinMax(*this, ICmpInst::ICMP_SLT, Ops);
11759   case X86::BI__builtin_ia32_pminub128:
11760   case X86::BI__builtin_ia32_pminuw128:
11761   case X86::BI__builtin_ia32_pminud128:
11762   case X86::BI__builtin_ia32_pminuq128:
11763   case X86::BI__builtin_ia32_pminub256:
11764   case X86::BI__builtin_ia32_pminuw256:
11765   case X86::BI__builtin_ia32_pminud256:
11766   case X86::BI__builtin_ia32_pminuq256:
11767   case X86::BI__builtin_ia32_pminub512:
11768   case X86::BI__builtin_ia32_pminuw512:
11769   case X86::BI__builtin_ia32_pminud512:
11770   case X86::BI__builtin_ia32_pminuq512:
11771     return EmitX86MinMax(*this, ICmpInst::ICMP_ULT, Ops);
11772 
11773   case X86::BI__builtin_ia32_pmuludq128:
11774   case X86::BI__builtin_ia32_pmuludq256:
11775   case X86::BI__builtin_ia32_pmuludq512:
11776     return EmitX86Muldq(*this, /*IsSigned*/false, Ops);
11777 
11778   case X86::BI__builtin_ia32_pmuldq128:
11779   case X86::BI__builtin_ia32_pmuldq256:
11780   case X86::BI__builtin_ia32_pmuldq512:
11781     return EmitX86Muldq(*this, /*IsSigned*/true, Ops);
11782 
11783   case X86::BI__builtin_ia32_pternlogd512_mask:
11784   case X86::BI__builtin_ia32_pternlogq512_mask:
11785   case X86::BI__builtin_ia32_pternlogd128_mask:
11786   case X86::BI__builtin_ia32_pternlogd256_mask:
11787   case X86::BI__builtin_ia32_pternlogq128_mask:
11788   case X86::BI__builtin_ia32_pternlogq256_mask:
11789     return EmitX86Ternlog(*this, /*ZeroMask*/false, Ops);
11790 
11791   case X86::BI__builtin_ia32_pternlogd512_maskz:
11792   case X86::BI__builtin_ia32_pternlogq512_maskz:
11793   case X86::BI__builtin_ia32_pternlogd128_maskz:
11794   case X86::BI__builtin_ia32_pternlogd256_maskz:
11795   case X86::BI__builtin_ia32_pternlogq128_maskz:
11796   case X86::BI__builtin_ia32_pternlogq256_maskz:
11797     return EmitX86Ternlog(*this, /*ZeroMask*/true, Ops);
11798 
11799   case X86::BI__builtin_ia32_vpshldd128:
11800   case X86::BI__builtin_ia32_vpshldd256:
11801   case X86::BI__builtin_ia32_vpshldd512:
11802   case X86::BI__builtin_ia32_vpshldq128:
11803   case X86::BI__builtin_ia32_vpshldq256:
11804   case X86::BI__builtin_ia32_vpshldq512:
11805   case X86::BI__builtin_ia32_vpshldw128:
11806   case X86::BI__builtin_ia32_vpshldw256:
11807   case X86::BI__builtin_ia32_vpshldw512:
11808     return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false);
11809 
11810   case X86::BI__builtin_ia32_vpshrdd128:
11811   case X86::BI__builtin_ia32_vpshrdd256:
11812   case X86::BI__builtin_ia32_vpshrdd512:
11813   case X86::BI__builtin_ia32_vpshrdq128:
11814   case X86::BI__builtin_ia32_vpshrdq256:
11815   case X86::BI__builtin_ia32_vpshrdq512:
11816   case X86::BI__builtin_ia32_vpshrdw128:
11817   case X86::BI__builtin_ia32_vpshrdw256:
11818   case X86::BI__builtin_ia32_vpshrdw512:
11819     // Ops 0 and 1 are swapped.
11820     return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true);
11821 
11822   case X86::BI__builtin_ia32_vpshldvd128:
11823   case X86::BI__builtin_ia32_vpshldvd256:
11824   case X86::BI__builtin_ia32_vpshldvd512:
11825   case X86::BI__builtin_ia32_vpshldvq128:
11826   case X86::BI__builtin_ia32_vpshldvq256:
11827   case X86::BI__builtin_ia32_vpshldvq512:
11828   case X86::BI__builtin_ia32_vpshldvw128:
11829   case X86::BI__builtin_ia32_vpshldvw256:
11830   case X86::BI__builtin_ia32_vpshldvw512:
11831     return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false);
11832 
11833   case X86::BI__builtin_ia32_vpshrdvd128:
11834   case X86::BI__builtin_ia32_vpshrdvd256:
11835   case X86::BI__builtin_ia32_vpshrdvd512:
11836   case X86::BI__builtin_ia32_vpshrdvq128:
11837   case X86::BI__builtin_ia32_vpshrdvq256:
11838   case X86::BI__builtin_ia32_vpshrdvq512:
11839   case X86::BI__builtin_ia32_vpshrdvw128:
11840   case X86::BI__builtin_ia32_vpshrdvw256:
11841   case X86::BI__builtin_ia32_vpshrdvw512:
11842     // Ops 0 and 1 are swapped.
11843     return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true);
11844 
11845   // 3DNow!
11846   case X86::BI__builtin_ia32_pswapdsf:
11847   case X86::BI__builtin_ia32_pswapdsi: {
11848     llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
11849     Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
11850     llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd);
11851     return Builder.CreateCall(F, Ops, "pswapd");
11852   }
11853   case X86::BI__builtin_ia32_rdrand16_step:
11854   case X86::BI__builtin_ia32_rdrand32_step:
11855   case X86::BI__builtin_ia32_rdrand64_step:
11856   case X86::BI__builtin_ia32_rdseed16_step:
11857   case X86::BI__builtin_ia32_rdseed32_step:
11858   case X86::BI__builtin_ia32_rdseed64_step: {
11859     Intrinsic::ID ID;
11860     switch (BuiltinID) {
11861     default: llvm_unreachable("Unsupported intrinsic!");
11862     case X86::BI__builtin_ia32_rdrand16_step:
11863       ID = Intrinsic::x86_rdrand_16;
11864       break;
11865     case X86::BI__builtin_ia32_rdrand32_step:
11866       ID = Intrinsic::x86_rdrand_32;
11867       break;
11868     case X86::BI__builtin_ia32_rdrand64_step:
11869       ID = Intrinsic::x86_rdrand_64;
11870       break;
11871     case X86::BI__builtin_ia32_rdseed16_step:
11872       ID = Intrinsic::x86_rdseed_16;
11873       break;
11874     case X86::BI__builtin_ia32_rdseed32_step:
11875       ID = Intrinsic::x86_rdseed_32;
11876       break;
11877     case X86::BI__builtin_ia32_rdseed64_step:
11878       ID = Intrinsic::x86_rdseed_64;
11879       break;
11880     }
11881 
11882     Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
11883     Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0),
11884                                       Ops[0]);
11885     return Builder.CreateExtractValue(Call, 1);
11886   }
11887   case X86::BI__builtin_ia32_addcarryx_u32:
11888   case X86::BI__builtin_ia32_addcarryx_u64:
11889   case X86::BI__builtin_ia32_subborrow_u32:
11890   case X86::BI__builtin_ia32_subborrow_u64: {
11891     Intrinsic::ID IID;
11892     switch (BuiltinID) {
11893     default: llvm_unreachable("Unsupported intrinsic!");
11894     case X86::BI__builtin_ia32_addcarryx_u32:
11895       IID = Intrinsic::x86_addcarry_32;
11896       break;
11897     case X86::BI__builtin_ia32_addcarryx_u64:
11898       IID = Intrinsic::x86_addcarry_64;
11899       break;
11900     case X86::BI__builtin_ia32_subborrow_u32:
11901       IID = Intrinsic::x86_subborrow_32;
11902       break;
11903     case X86::BI__builtin_ia32_subborrow_u64:
11904       IID = Intrinsic::x86_subborrow_64;
11905       break;
11906     }
11907 
11908     Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID),
11909                                      { Ops[0], Ops[1], Ops[2] });
11910     Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1),
11911                                       Ops[3]);
11912     return Builder.CreateExtractValue(Call, 0);
11913   }
11914 
11915   case X86::BI__builtin_ia32_fpclassps128_mask:
11916   case X86::BI__builtin_ia32_fpclassps256_mask:
11917   case X86::BI__builtin_ia32_fpclassps512_mask:
11918   case X86::BI__builtin_ia32_fpclasspd128_mask:
11919   case X86::BI__builtin_ia32_fpclasspd256_mask:
11920   case X86::BI__builtin_ia32_fpclasspd512_mask: {
11921     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
11922     Value *MaskIn = Ops[2];
11923     Ops.erase(&Ops[2]);
11924 
11925     Intrinsic::ID ID;
11926     switch (BuiltinID) {
11927     default: llvm_unreachable("Unsupported intrinsic!");
11928     case X86::BI__builtin_ia32_fpclassps128_mask:
11929       ID = Intrinsic::x86_avx512_fpclass_ps_128;
11930       break;
11931     case X86::BI__builtin_ia32_fpclassps256_mask:
11932       ID = Intrinsic::x86_avx512_fpclass_ps_256;
11933       break;
11934     case X86::BI__builtin_ia32_fpclassps512_mask:
11935       ID = Intrinsic::x86_avx512_fpclass_ps_512;
11936       break;
11937     case X86::BI__builtin_ia32_fpclasspd128_mask:
11938       ID = Intrinsic::x86_avx512_fpclass_pd_128;
11939       break;
11940     case X86::BI__builtin_ia32_fpclasspd256_mask:
11941       ID = Intrinsic::x86_avx512_fpclass_pd_256;
11942       break;
11943     case X86::BI__builtin_ia32_fpclasspd512_mask:
11944       ID = Intrinsic::x86_avx512_fpclass_pd_512;
11945       break;
11946     }
11947 
11948     Value *Fpclass = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
11949     return EmitX86MaskedCompareResult(*this, Fpclass, NumElts, MaskIn);
11950   }
11951 
11952   case X86::BI__builtin_ia32_vp2intersect_q_512:
11953   case X86::BI__builtin_ia32_vp2intersect_q_256:
11954   case X86::BI__builtin_ia32_vp2intersect_q_128:
11955   case X86::BI__builtin_ia32_vp2intersect_d_512:
11956   case X86::BI__builtin_ia32_vp2intersect_d_256:
11957   case X86::BI__builtin_ia32_vp2intersect_d_128: {
11958     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
11959     Intrinsic::ID ID;
11960 
11961     switch (BuiltinID) {
11962     default: llvm_unreachable("Unsupported intrinsic!");
11963     case X86::BI__builtin_ia32_vp2intersect_q_512:
11964       ID = Intrinsic::x86_avx512_vp2intersect_q_512;
11965       break;
11966     case X86::BI__builtin_ia32_vp2intersect_q_256:
11967       ID = Intrinsic::x86_avx512_vp2intersect_q_256;
11968       break;
11969     case X86::BI__builtin_ia32_vp2intersect_q_128:
11970       ID = Intrinsic::x86_avx512_vp2intersect_q_128;
11971       break;
11972     case X86::BI__builtin_ia32_vp2intersect_d_512:
11973       ID = Intrinsic::x86_avx512_vp2intersect_d_512;
11974       break;
11975     case X86::BI__builtin_ia32_vp2intersect_d_256:
11976       ID = Intrinsic::x86_avx512_vp2intersect_d_256;
11977       break;
11978     case X86::BI__builtin_ia32_vp2intersect_d_128:
11979       ID = Intrinsic::x86_avx512_vp2intersect_d_128;
11980       break;
11981     }
11982 
11983     Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {Ops[0], Ops[1]});
11984     Value *Result = Builder.CreateExtractValue(Call, 0);
11985     Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr);
11986     Builder.CreateDefaultAlignedStore(Result, Ops[2]);
11987 
11988     Result = Builder.CreateExtractValue(Call, 1);
11989     Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr);
11990     return Builder.CreateDefaultAlignedStore(Result, Ops[3]);
11991   }
11992 
11993   case X86::BI__builtin_ia32_vpmultishiftqb128:
11994   case X86::BI__builtin_ia32_vpmultishiftqb256:
11995   case X86::BI__builtin_ia32_vpmultishiftqb512: {
11996     Intrinsic::ID ID;
11997     switch (BuiltinID) {
11998     default: llvm_unreachable("Unsupported intrinsic!");
11999     case X86::BI__builtin_ia32_vpmultishiftqb128:
12000       ID = Intrinsic::x86_avx512_pmultishift_qb_128;
12001       break;
12002     case X86::BI__builtin_ia32_vpmultishiftqb256:
12003       ID = Intrinsic::x86_avx512_pmultishift_qb_256;
12004       break;
12005     case X86::BI__builtin_ia32_vpmultishiftqb512:
12006       ID = Intrinsic::x86_avx512_pmultishift_qb_512;
12007       break;
12008     }
12009 
12010     return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
12011   }
12012 
12013   case X86::BI__builtin_ia32_vpshufbitqmb128_mask:
12014   case X86::BI__builtin_ia32_vpshufbitqmb256_mask:
12015   case X86::BI__builtin_ia32_vpshufbitqmb512_mask: {
12016     unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
12017     Value *MaskIn = Ops[2];
12018     Ops.erase(&Ops[2]);
12019 
12020     Intrinsic::ID ID;
12021     switch (BuiltinID) {
12022     default: llvm_unreachable("Unsupported intrinsic!");
12023     case X86::BI__builtin_ia32_vpshufbitqmb128_mask:
12024       ID = Intrinsic::x86_avx512_vpshufbitqmb_128;
12025       break;
12026     case X86::BI__builtin_ia32_vpshufbitqmb256_mask:
12027       ID = Intrinsic::x86_avx512_vpshufbitqmb_256;
12028       break;
12029     case X86::BI__builtin_ia32_vpshufbitqmb512_mask:
12030       ID = Intrinsic::x86_avx512_vpshufbitqmb_512;
12031       break;
12032     }
12033 
12034     Value *Shufbit = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
12035     return EmitX86MaskedCompareResult(*this, Shufbit, NumElts, MaskIn);
12036   }
12037 
12038   // packed comparison intrinsics
12039   case X86::BI__builtin_ia32_cmpeqps:
12040   case X86::BI__builtin_ia32_cmpeqpd:
12041     return getVectorFCmpIR(CmpInst::FCMP_OEQ);
12042   case X86::BI__builtin_ia32_cmpltps:
12043   case X86::BI__builtin_ia32_cmpltpd:
12044     return getVectorFCmpIR(CmpInst::FCMP_OLT);
12045   case X86::BI__builtin_ia32_cmpleps:
12046   case X86::BI__builtin_ia32_cmplepd:
12047     return getVectorFCmpIR(CmpInst::FCMP_OLE);
12048   case X86::BI__builtin_ia32_cmpunordps:
12049   case X86::BI__builtin_ia32_cmpunordpd:
12050     return getVectorFCmpIR(CmpInst::FCMP_UNO);
12051   case X86::BI__builtin_ia32_cmpneqps:
12052   case X86::BI__builtin_ia32_cmpneqpd:
12053     return getVectorFCmpIR(CmpInst::FCMP_UNE);
12054   case X86::BI__builtin_ia32_cmpnltps:
12055   case X86::BI__builtin_ia32_cmpnltpd:
12056     return getVectorFCmpIR(CmpInst::FCMP_UGE);
12057   case X86::BI__builtin_ia32_cmpnleps:
12058   case X86::BI__builtin_ia32_cmpnlepd:
12059     return getVectorFCmpIR(CmpInst::FCMP_UGT);
12060   case X86::BI__builtin_ia32_cmpordps:
12061   case X86::BI__builtin_ia32_cmpordpd:
12062     return getVectorFCmpIR(CmpInst::FCMP_ORD);
12063   case X86::BI__builtin_ia32_cmpps:
12064   case X86::BI__builtin_ia32_cmpps256:
12065   case X86::BI__builtin_ia32_cmppd:
12066   case X86::BI__builtin_ia32_cmppd256:
12067   case X86::BI__builtin_ia32_cmpps128_mask:
12068   case X86::BI__builtin_ia32_cmpps256_mask:
12069   case X86::BI__builtin_ia32_cmpps512_mask:
12070   case X86::BI__builtin_ia32_cmppd128_mask:
12071   case X86::BI__builtin_ia32_cmppd256_mask:
12072   case X86::BI__builtin_ia32_cmppd512_mask: {
12073     // Lowering vector comparisons to fcmp instructions, while
12074     // ignoring signalling behaviour requested
12075     // ignoring rounding mode requested
12076     // This is is only possible as long as FENV_ACCESS is not implemented.
12077     // See also: https://reviews.llvm.org/D45616
12078 
12079     // The third argument is the comparison condition, and integer in the
12080     // range [0, 31]
12081     unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x1f;
12082 
12083     // Lowering to IR fcmp instruction.
12084     // Ignoring requested signaling behaviour,
12085     // e.g. both _CMP_GT_OS & _CMP_GT_OQ are translated to FCMP_OGT.
12086     FCmpInst::Predicate Pred;
12087     switch (CC) {
12088     case 0x00: Pred = FCmpInst::FCMP_OEQ;   break;
12089     case 0x01: Pred = FCmpInst::FCMP_OLT;   break;
12090     case 0x02: Pred = FCmpInst::FCMP_OLE;   break;
12091     case 0x03: Pred = FCmpInst::FCMP_UNO;   break;
12092     case 0x04: Pred = FCmpInst::FCMP_UNE;   break;
12093     case 0x05: Pred = FCmpInst::FCMP_UGE;   break;
12094     case 0x06: Pred = FCmpInst::FCMP_UGT;   break;
12095     case 0x07: Pred = FCmpInst::FCMP_ORD;   break;
12096     case 0x08: Pred = FCmpInst::FCMP_UEQ;   break;
12097     case 0x09: Pred = FCmpInst::FCMP_ULT;   break;
12098     case 0x0a: Pred = FCmpInst::FCMP_ULE;   break;
12099     case 0x0b: Pred = FCmpInst::FCMP_FALSE; break;
12100     case 0x0c: Pred = FCmpInst::FCMP_ONE;   break;
12101     case 0x0d: Pred = FCmpInst::FCMP_OGE;   break;
12102     case 0x0e: Pred = FCmpInst::FCMP_OGT;   break;
12103     case 0x0f: Pred = FCmpInst::FCMP_TRUE;  break;
12104     case 0x10: Pred = FCmpInst::FCMP_OEQ;   break;
12105     case 0x11: Pred = FCmpInst::FCMP_OLT;   break;
12106     case 0x12: Pred = FCmpInst::FCMP_OLE;   break;
12107     case 0x13: Pred = FCmpInst::FCMP_UNO;   break;
12108     case 0x14: Pred = FCmpInst::FCMP_UNE;   break;
12109     case 0x15: Pred = FCmpInst::FCMP_UGE;   break;
12110     case 0x16: Pred = FCmpInst::FCMP_UGT;   break;
12111     case 0x17: Pred = FCmpInst::FCMP_ORD;   break;
12112     case 0x18: Pred = FCmpInst::FCMP_UEQ;   break;
12113     case 0x19: Pred = FCmpInst::FCMP_ULT;   break;
12114     case 0x1a: Pred = FCmpInst::FCMP_ULE;   break;
12115     case 0x1b: Pred = FCmpInst::FCMP_FALSE; break;
12116     case 0x1c: Pred = FCmpInst::FCMP_ONE;   break;
12117     case 0x1d: Pred = FCmpInst::FCMP_OGE;   break;
12118     case 0x1e: Pred = FCmpInst::FCMP_OGT;   break;
12119     case 0x1f: Pred = FCmpInst::FCMP_TRUE;  break;
12120     default: llvm_unreachable("Unhandled CC");
12121     }
12122 
12123     // Builtins without the _mask suffix return a vector of integers
12124     // of the same width as the input vectors
12125     switch (BuiltinID) {
12126     case X86::BI__builtin_ia32_cmpps512_mask:
12127     case X86::BI__builtin_ia32_cmppd512_mask:
12128     case X86::BI__builtin_ia32_cmpps128_mask:
12129     case X86::BI__builtin_ia32_cmpps256_mask:
12130     case X86::BI__builtin_ia32_cmppd128_mask:
12131     case X86::BI__builtin_ia32_cmppd256_mask: {
12132       unsigned NumElts = Ops[0]->getType()->getVectorNumElements();
12133       Value *Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]);
12134       return EmitX86MaskedCompareResult(*this, Cmp, NumElts, Ops[3]);
12135     }
12136     default:
12137       return getVectorFCmpIR(Pred);
12138     }
12139   }
12140 
12141   // SSE scalar comparison intrinsics
12142   case X86::BI__builtin_ia32_cmpeqss:
12143     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 0);
12144   case X86::BI__builtin_ia32_cmpltss:
12145     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 1);
12146   case X86::BI__builtin_ia32_cmpless:
12147     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 2);
12148   case X86::BI__builtin_ia32_cmpunordss:
12149     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 3);
12150   case X86::BI__builtin_ia32_cmpneqss:
12151     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 4);
12152   case X86::BI__builtin_ia32_cmpnltss:
12153     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 5);
12154   case X86::BI__builtin_ia32_cmpnless:
12155     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 6);
12156   case X86::BI__builtin_ia32_cmpordss:
12157     return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 7);
12158   case X86::BI__builtin_ia32_cmpeqsd:
12159     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 0);
12160   case X86::BI__builtin_ia32_cmpltsd:
12161     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 1);
12162   case X86::BI__builtin_ia32_cmplesd:
12163     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 2);
12164   case X86::BI__builtin_ia32_cmpunordsd:
12165     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 3);
12166   case X86::BI__builtin_ia32_cmpneqsd:
12167     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4);
12168   case X86::BI__builtin_ia32_cmpnltsd:
12169     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5);
12170   case X86::BI__builtin_ia32_cmpnlesd:
12171     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6);
12172   case X86::BI__builtin_ia32_cmpordsd:
12173     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7);
12174 
12175 // AVX512 bf16 intrinsics
12176   case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: {
12177     Ops[2] = getMaskVecValue(*this, Ops[2],
12178                              Ops[0]->getType()->getVectorNumElements());
12179     Intrinsic::ID IID = Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128;
12180     return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
12181   }
12182   case X86::BI__builtin_ia32_cvtsbf162ss_32:
12183     return EmitX86CvtBF16ToFloatExpr(*this, E, Ops);
12184 
12185   case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
12186   case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: {
12187     Intrinsic::ID IID;
12188     switch (BuiltinID) {
12189     default: llvm_unreachable("Unsupported intrinsic!");
12190     case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
12191       IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_256;
12192       break;
12193     case X86::BI__builtin_ia32_cvtneps2bf16_512_mask:
12194       IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_512;
12195       break;
12196     }
12197     Value *Res = Builder.CreateCall(CGM.getIntrinsic(IID), Ops[0]);
12198     return EmitX86Select(*this, Ops[2], Res, Ops[1]);
12199   }
12200 
12201   case X86::BI__emul:
12202   case X86::BI__emulu: {
12203     llvm::Type *Int64Ty = llvm::IntegerType::get(getLLVMContext(), 64);
12204     bool isSigned = (BuiltinID == X86::BI__emul);
12205     Value *LHS = Builder.CreateIntCast(Ops[0], Int64Ty, isSigned);
12206     Value *RHS = Builder.CreateIntCast(Ops[1], Int64Ty, isSigned);
12207     return Builder.CreateMul(LHS, RHS, "", !isSigned, isSigned);
12208   }
12209   case X86::BI__mulh:
12210   case X86::BI__umulh:
12211   case X86::BI_mul128:
12212   case X86::BI_umul128: {
12213     llvm::Type *ResType = ConvertType(E->getType());
12214     llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
12215 
12216     bool IsSigned = (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI_mul128);
12217     Value *LHS = Builder.CreateIntCast(Ops[0], Int128Ty, IsSigned);
12218     Value *RHS = Builder.CreateIntCast(Ops[1], Int128Ty, IsSigned);
12219 
12220     Value *MulResult, *HigherBits;
12221     if (IsSigned) {
12222       MulResult = Builder.CreateNSWMul(LHS, RHS);
12223       HigherBits = Builder.CreateAShr(MulResult, 64);
12224     } else {
12225       MulResult = Builder.CreateNUWMul(LHS, RHS);
12226       HigherBits = Builder.CreateLShr(MulResult, 64);
12227     }
12228     HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned);
12229 
12230     if (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI__umulh)
12231       return HigherBits;
12232 
12233     Address HighBitsAddress = EmitPointerWithAlignment(E->getArg(2));
12234     Builder.CreateStore(HigherBits, HighBitsAddress);
12235     return Builder.CreateIntCast(MulResult, ResType, IsSigned);
12236   }
12237 
12238   case X86::BI__faststorefence: {
12239     return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
12240                                llvm::SyncScope::System);
12241   }
12242   case X86::BI__shiftleft128:
12243   case X86::BI__shiftright128: {
12244     // FIXME: Once fshl/fshr no longer add an unneeded and and cmov, do this:
12245     // llvm::Function *F = CGM.getIntrinsic(
12246     //   BuiltinID == X86::BI__shiftleft128 ? Intrinsic::fshl : Intrinsic::fshr,
12247     //   Int64Ty);
12248     // Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
12249     // return Builder.CreateCall(F, Ops);
12250     llvm::Type *Int128Ty = Builder.getInt128Ty();
12251     Value *HighPart128 =
12252         Builder.CreateShl(Builder.CreateZExt(Ops[1], Int128Ty), 64);
12253     Value *LowPart128 = Builder.CreateZExt(Ops[0], Int128Ty);
12254     Value *Val = Builder.CreateOr(HighPart128, LowPart128);
12255     Value *Amt = Builder.CreateAnd(Builder.CreateZExt(Ops[2], Int128Ty),
12256                                    llvm::ConstantInt::get(Int128Ty, 0x3f));
12257     Value *Res;
12258     if (BuiltinID == X86::BI__shiftleft128)
12259       Res = Builder.CreateLShr(Builder.CreateShl(Val, Amt), 64);
12260     else
12261       Res = Builder.CreateLShr(Val, Amt);
12262     return Builder.CreateTrunc(Res, Int64Ty);
12263   }
12264   case X86::BI_ReadWriteBarrier:
12265   case X86::BI_ReadBarrier:
12266   case X86::BI_WriteBarrier: {
12267     return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
12268                                llvm::SyncScope::SingleThread);
12269   }
12270   case X86::BI_BitScanForward:
12271   case X86::BI_BitScanForward64:
12272     return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E);
12273   case X86::BI_BitScanReverse:
12274   case X86::BI_BitScanReverse64:
12275     return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E);
12276 
12277   case X86::BI_InterlockedAnd64:
12278     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E);
12279   case X86::BI_InterlockedExchange64:
12280     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E);
12281   case X86::BI_InterlockedExchangeAdd64:
12282     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E);
12283   case X86::BI_InterlockedExchangeSub64:
12284     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E);
12285   case X86::BI_InterlockedOr64:
12286     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E);
12287   case X86::BI_InterlockedXor64:
12288     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E);
12289   case X86::BI_InterlockedDecrement64:
12290     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E);
12291   case X86::BI_InterlockedIncrement64:
12292     return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E);
12293   case X86::BI_InterlockedCompareExchange128: {
12294     // InterlockedCompareExchange128 doesn't directly refer to 128bit ints,
12295     // instead it takes pointers to 64bit ints for Destination and
12296     // ComparandResult, and exchange is taken as two 64bit ints (high & low).
12297     // The previous value is written to ComparandResult, and success is
12298     // returned.
12299 
12300     llvm::Type *Int128Ty = Builder.getInt128Ty();
12301     llvm::Type *Int128PtrTy = Int128Ty->getPointerTo();
12302 
12303     Value *Destination =
12304         Builder.CreateBitCast(Ops[0], Int128PtrTy);
12305     Value *ExchangeHigh128 = Builder.CreateZExt(Ops[1], Int128Ty);
12306     Value *ExchangeLow128 = Builder.CreateZExt(Ops[2], Int128Ty);
12307     Address ComparandResult(Builder.CreateBitCast(Ops[3], Int128PtrTy),
12308                             getContext().toCharUnitsFromBits(128));
12309 
12310     Value *Exchange = Builder.CreateOr(
12311         Builder.CreateShl(ExchangeHigh128, 64, "", false, false),
12312         ExchangeLow128);
12313 
12314     Value *Comparand = Builder.CreateLoad(ComparandResult);
12315 
12316     AtomicCmpXchgInst *CXI =
12317         Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
12318                                     AtomicOrdering::SequentiallyConsistent,
12319                                     AtomicOrdering::SequentiallyConsistent);
12320     CXI->setVolatile(true);
12321 
12322     // Write the result back to the inout pointer.
12323     Builder.CreateStore(Builder.CreateExtractValue(CXI, 0), ComparandResult);
12324 
12325     // Get the success boolean and zero extend it to i8.
12326     Value *Success = Builder.CreateExtractValue(CXI, 1);
12327     return Builder.CreateZExt(Success, ConvertType(E->getType()));
12328   }
12329 
12330   case X86::BI_AddressOfReturnAddress: {
12331     Function *F =
12332         CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy);
12333     return Builder.CreateCall(F);
12334   }
12335   case X86::BI__stosb: {
12336     // We treat __stosb as a volatile memset - it may not generate "rep stosb"
12337     // instruction, but it will create a memset that won't be optimized away.
12338     return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], 1, true);
12339   }
12340   case X86::BI__ud2:
12341     // llvm.trap makes a ud2a instruction on x86.
12342     return EmitTrapCall(Intrinsic::trap);
12343   case X86::BI__int2c: {
12344     // This syscall signals a driver assertion failure in x86 NT kernels.
12345     llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
12346     llvm::InlineAsm *IA =
12347         llvm::InlineAsm::get(FTy, "int $$0x2c", "", /*hasSideEffects=*/true);
12348     llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
12349         getLLVMContext(), llvm::AttributeList::FunctionIndex,
12350         llvm::Attribute::NoReturn);
12351     llvm::CallInst *CI = Builder.CreateCall(IA);
12352     CI->setAttributes(NoReturnAttr);
12353     return CI;
12354   }
12355   case X86::BI__readfsbyte:
12356   case X86::BI__readfsword:
12357   case X86::BI__readfsdword:
12358   case X86::BI__readfsqword: {
12359     llvm::Type *IntTy = ConvertType(E->getType());
12360     Value *Ptr =
12361         Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 257));
12362     LoadInst *Load = Builder.CreateAlignedLoad(
12363         IntTy, Ptr, getContext().getTypeAlignInChars(E->getType()));
12364     Load->setVolatile(true);
12365     return Load;
12366   }
12367   case X86::BI__readgsbyte:
12368   case X86::BI__readgsword:
12369   case X86::BI__readgsdword:
12370   case X86::BI__readgsqword: {
12371     llvm::Type *IntTy = ConvertType(E->getType());
12372     Value *Ptr =
12373         Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 256));
12374     LoadInst *Load = Builder.CreateAlignedLoad(
12375         IntTy, Ptr, getContext().getTypeAlignInChars(E->getType()));
12376     Load->setVolatile(true);
12377     return Load;
12378   }
12379   case X86::BI__builtin_ia32_paddsb512:
12380   case X86::BI__builtin_ia32_paddsw512:
12381   case X86::BI__builtin_ia32_paddsb256:
12382   case X86::BI__builtin_ia32_paddsw256:
12383   case X86::BI__builtin_ia32_paddsb128:
12384   case X86::BI__builtin_ia32_paddsw128:
12385     return EmitX86AddSubSatExpr(*this, Ops, true, true);
12386   case X86::BI__builtin_ia32_paddusb512:
12387   case X86::BI__builtin_ia32_paddusw512:
12388   case X86::BI__builtin_ia32_paddusb256:
12389   case X86::BI__builtin_ia32_paddusw256:
12390   case X86::BI__builtin_ia32_paddusb128:
12391   case X86::BI__builtin_ia32_paddusw128:
12392     return EmitX86AddSubSatExpr(*this, Ops, false, true);
12393   case X86::BI__builtin_ia32_psubsb512:
12394   case X86::BI__builtin_ia32_psubsw512:
12395   case X86::BI__builtin_ia32_psubsb256:
12396   case X86::BI__builtin_ia32_psubsw256:
12397   case X86::BI__builtin_ia32_psubsb128:
12398   case X86::BI__builtin_ia32_psubsw128:
12399     return EmitX86AddSubSatExpr(*this, Ops, true, false);
12400   case X86::BI__builtin_ia32_psubusb512:
12401   case X86::BI__builtin_ia32_psubusw512:
12402   case X86::BI__builtin_ia32_psubusb256:
12403   case X86::BI__builtin_ia32_psubusw256:
12404   case X86::BI__builtin_ia32_psubusb128:
12405   case X86::BI__builtin_ia32_psubusw128:
12406     return EmitX86AddSubSatExpr(*this, Ops, false, false);
12407   }
12408 }
12409 
12410 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
12411                                            const CallExpr *E) {
12412   SmallVector<Value*, 4> Ops;
12413 
12414   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
12415     Ops.push_back(EmitScalarExpr(E->getArg(i)));
12416 
12417   Intrinsic::ID ID = Intrinsic::not_intrinsic;
12418 
12419   switch (BuiltinID) {
12420   default: return nullptr;
12421 
12422   // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we
12423   // call __builtin_readcyclecounter.
12424   case PPC::BI__builtin_ppc_get_timebase:
12425     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter));
12426 
12427   // vec_ld, vec_xl_be, vec_lvsl, vec_lvsr
12428   case PPC::BI__builtin_altivec_lvx:
12429   case PPC::BI__builtin_altivec_lvxl:
12430   case PPC::BI__builtin_altivec_lvebx:
12431   case PPC::BI__builtin_altivec_lvehx:
12432   case PPC::BI__builtin_altivec_lvewx:
12433   case PPC::BI__builtin_altivec_lvsl:
12434   case PPC::BI__builtin_altivec_lvsr:
12435   case PPC::BI__builtin_vsx_lxvd2x:
12436   case PPC::BI__builtin_vsx_lxvw4x:
12437   case PPC::BI__builtin_vsx_lxvd2x_be:
12438   case PPC::BI__builtin_vsx_lxvw4x_be:
12439   case PPC::BI__builtin_vsx_lxvl:
12440   case PPC::BI__builtin_vsx_lxvll:
12441   {
12442     if(BuiltinID == PPC::BI__builtin_vsx_lxvl ||
12443        BuiltinID == PPC::BI__builtin_vsx_lxvll){
12444       Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy);
12445     }else {
12446       Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
12447       Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]);
12448       Ops.pop_back();
12449     }
12450 
12451     switch (BuiltinID) {
12452     default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
12453     case PPC::BI__builtin_altivec_lvx:
12454       ID = Intrinsic::ppc_altivec_lvx;
12455       break;
12456     case PPC::BI__builtin_altivec_lvxl:
12457       ID = Intrinsic::ppc_altivec_lvxl;
12458       break;
12459     case PPC::BI__builtin_altivec_lvebx:
12460       ID = Intrinsic::ppc_altivec_lvebx;
12461       break;
12462     case PPC::BI__builtin_altivec_lvehx:
12463       ID = Intrinsic::ppc_altivec_lvehx;
12464       break;
12465     case PPC::BI__builtin_altivec_lvewx:
12466       ID = Intrinsic::ppc_altivec_lvewx;
12467       break;
12468     case PPC::BI__builtin_altivec_lvsl:
12469       ID = Intrinsic::ppc_altivec_lvsl;
12470       break;
12471     case PPC::BI__builtin_altivec_lvsr:
12472       ID = Intrinsic::ppc_altivec_lvsr;
12473       break;
12474     case PPC::BI__builtin_vsx_lxvd2x:
12475       ID = Intrinsic::ppc_vsx_lxvd2x;
12476       break;
12477     case PPC::BI__builtin_vsx_lxvw4x:
12478       ID = Intrinsic::ppc_vsx_lxvw4x;
12479       break;
12480     case PPC::BI__builtin_vsx_lxvd2x_be:
12481       ID = Intrinsic::ppc_vsx_lxvd2x_be;
12482       break;
12483     case PPC::BI__builtin_vsx_lxvw4x_be:
12484       ID = Intrinsic::ppc_vsx_lxvw4x_be;
12485       break;
12486     case PPC::BI__builtin_vsx_lxvl:
12487       ID = Intrinsic::ppc_vsx_lxvl;
12488       break;
12489     case PPC::BI__builtin_vsx_lxvll:
12490       ID = Intrinsic::ppc_vsx_lxvll;
12491       break;
12492     }
12493     llvm::Function *F = CGM.getIntrinsic(ID);
12494     return Builder.CreateCall(F, Ops, "");
12495   }
12496 
12497   // vec_st, vec_xst_be
12498   case PPC::BI__builtin_altivec_stvx:
12499   case PPC::BI__builtin_altivec_stvxl:
12500   case PPC::BI__builtin_altivec_stvebx:
12501   case PPC::BI__builtin_altivec_stvehx:
12502   case PPC::BI__builtin_altivec_stvewx:
12503   case PPC::BI__builtin_vsx_stxvd2x:
12504   case PPC::BI__builtin_vsx_stxvw4x:
12505   case PPC::BI__builtin_vsx_stxvd2x_be:
12506   case PPC::BI__builtin_vsx_stxvw4x_be:
12507   case PPC::BI__builtin_vsx_stxvl:
12508   case PPC::BI__builtin_vsx_stxvll:
12509   {
12510     if(BuiltinID == PPC::BI__builtin_vsx_stxvl ||
12511       BuiltinID == PPC::BI__builtin_vsx_stxvll ){
12512       Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
12513     }else {
12514       Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
12515       Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]);
12516       Ops.pop_back();
12517     }
12518 
12519     switch (BuiltinID) {
12520     default: llvm_unreachable("Unsupported st intrinsic!");
12521     case PPC::BI__builtin_altivec_stvx:
12522       ID = Intrinsic::ppc_altivec_stvx;
12523       break;
12524     case PPC::BI__builtin_altivec_stvxl:
12525       ID = Intrinsic::ppc_altivec_stvxl;
12526       break;
12527     case PPC::BI__builtin_altivec_stvebx:
12528       ID = Intrinsic::ppc_altivec_stvebx;
12529       break;
12530     case PPC::BI__builtin_altivec_stvehx:
12531       ID = Intrinsic::ppc_altivec_stvehx;
12532       break;
12533     case PPC::BI__builtin_altivec_stvewx:
12534       ID = Intrinsic::ppc_altivec_stvewx;
12535       break;
12536     case PPC::BI__builtin_vsx_stxvd2x:
12537       ID = Intrinsic::ppc_vsx_stxvd2x;
12538       break;
12539     case PPC::BI__builtin_vsx_stxvw4x:
12540       ID = Intrinsic::ppc_vsx_stxvw4x;
12541       break;
12542     case PPC::BI__builtin_vsx_stxvd2x_be:
12543       ID = Intrinsic::ppc_vsx_stxvd2x_be;
12544       break;
12545     case PPC::BI__builtin_vsx_stxvw4x_be:
12546       ID = Intrinsic::ppc_vsx_stxvw4x_be;
12547       break;
12548     case PPC::BI__builtin_vsx_stxvl:
12549       ID = Intrinsic::ppc_vsx_stxvl;
12550       break;
12551     case PPC::BI__builtin_vsx_stxvll:
12552       ID = Intrinsic::ppc_vsx_stxvll;
12553       break;
12554     }
12555     llvm::Function *F = CGM.getIntrinsic(ID);
12556     return Builder.CreateCall(F, Ops, "");
12557   }
12558   // Square root
12559   case PPC::BI__builtin_vsx_xvsqrtsp:
12560   case PPC::BI__builtin_vsx_xvsqrtdp: {
12561     llvm::Type *ResultType = ConvertType(E->getType());
12562     Value *X = EmitScalarExpr(E->getArg(0));
12563     ID = Intrinsic::sqrt;
12564     llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
12565     return Builder.CreateCall(F, X);
12566   }
12567   // Count leading zeros
12568   case PPC::BI__builtin_altivec_vclzb:
12569   case PPC::BI__builtin_altivec_vclzh:
12570   case PPC::BI__builtin_altivec_vclzw:
12571   case PPC::BI__builtin_altivec_vclzd: {
12572     llvm::Type *ResultType = ConvertType(E->getType());
12573     Value *X = EmitScalarExpr(E->getArg(0));
12574     Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
12575     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType);
12576     return Builder.CreateCall(F, {X, Undef});
12577   }
12578   case PPC::BI__builtin_altivec_vctzb:
12579   case PPC::BI__builtin_altivec_vctzh:
12580   case PPC::BI__builtin_altivec_vctzw:
12581   case PPC::BI__builtin_altivec_vctzd: {
12582     llvm::Type *ResultType = ConvertType(E->getType());
12583     Value *X = EmitScalarExpr(E->getArg(0));
12584     Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
12585     Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType);
12586     return Builder.CreateCall(F, {X, Undef});
12587   }
12588   case PPC::BI__builtin_altivec_vpopcntb:
12589   case PPC::BI__builtin_altivec_vpopcnth:
12590   case PPC::BI__builtin_altivec_vpopcntw:
12591   case PPC::BI__builtin_altivec_vpopcntd: {
12592     llvm::Type *ResultType = ConvertType(E->getType());
12593     Value *X = EmitScalarExpr(E->getArg(0));
12594     llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
12595     return Builder.CreateCall(F, X);
12596   }
12597   // Copy sign
12598   case PPC::BI__builtin_vsx_xvcpsgnsp:
12599   case PPC::BI__builtin_vsx_xvcpsgndp: {
12600     llvm::Type *ResultType = ConvertType(E->getType());
12601     Value *X = EmitScalarExpr(E->getArg(0));
12602     Value *Y = EmitScalarExpr(E->getArg(1));
12603     ID = Intrinsic::copysign;
12604     llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
12605     return Builder.CreateCall(F, {X, Y});
12606   }
12607   // Rounding/truncation
12608   case PPC::BI__builtin_vsx_xvrspip:
12609   case PPC::BI__builtin_vsx_xvrdpip:
12610   case PPC::BI__builtin_vsx_xvrdpim:
12611   case PPC::BI__builtin_vsx_xvrspim:
12612   case PPC::BI__builtin_vsx_xvrdpi:
12613   case PPC::BI__builtin_vsx_xvrspi:
12614   case PPC::BI__builtin_vsx_xvrdpic:
12615   case PPC::BI__builtin_vsx_xvrspic:
12616   case PPC::BI__builtin_vsx_xvrdpiz:
12617   case PPC::BI__builtin_vsx_xvrspiz: {
12618     llvm::Type *ResultType = ConvertType(E->getType());
12619     Value *X = EmitScalarExpr(E->getArg(0));
12620     if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim ||
12621         BuiltinID == PPC::BI__builtin_vsx_xvrspim)
12622       ID = Intrinsic::floor;
12623     else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi ||
12624              BuiltinID == PPC::BI__builtin_vsx_xvrspi)
12625       ID = Intrinsic::round;
12626     else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic ||
12627              BuiltinID == PPC::BI__builtin_vsx_xvrspic)
12628       ID = Intrinsic::nearbyint;
12629     else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip ||
12630              BuiltinID == PPC::BI__builtin_vsx_xvrspip)
12631       ID = Intrinsic::ceil;
12632     else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz ||
12633              BuiltinID == PPC::BI__builtin_vsx_xvrspiz)
12634       ID = Intrinsic::trunc;
12635     llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
12636     return Builder.CreateCall(F, X);
12637   }
12638 
12639   // Absolute value
12640   case PPC::BI__builtin_vsx_xvabsdp:
12641   case PPC::BI__builtin_vsx_xvabssp: {
12642     llvm::Type *ResultType = ConvertType(E->getType());
12643     Value *X = EmitScalarExpr(E->getArg(0));
12644     llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
12645     return Builder.CreateCall(F, X);
12646   }
12647 
12648   // FMA variations
12649   case PPC::BI__builtin_vsx_xvmaddadp:
12650   case PPC::BI__builtin_vsx_xvmaddasp:
12651   case PPC::BI__builtin_vsx_xvnmaddadp:
12652   case PPC::BI__builtin_vsx_xvnmaddasp:
12653   case PPC::BI__builtin_vsx_xvmsubadp:
12654   case PPC::BI__builtin_vsx_xvmsubasp:
12655   case PPC::BI__builtin_vsx_xvnmsubadp:
12656   case PPC::BI__builtin_vsx_xvnmsubasp: {
12657     llvm::Type *ResultType = ConvertType(E->getType());
12658     Value *X = EmitScalarExpr(E->getArg(0));
12659     Value *Y = EmitScalarExpr(E->getArg(1));
12660     Value *Z = EmitScalarExpr(E->getArg(2));
12661     Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType);
12662     llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
12663     switch (BuiltinID) {
12664       case PPC::BI__builtin_vsx_xvmaddadp:
12665       case PPC::BI__builtin_vsx_xvmaddasp:
12666         return Builder.CreateCall(F, {X, Y, Z});
12667       case PPC::BI__builtin_vsx_xvnmaddadp:
12668       case PPC::BI__builtin_vsx_xvnmaddasp:
12669         return Builder.CreateFSub(Zero,
12670                                   Builder.CreateCall(F, {X, Y, Z}), "sub");
12671       case PPC::BI__builtin_vsx_xvmsubadp:
12672       case PPC::BI__builtin_vsx_xvmsubasp:
12673         return Builder.CreateCall(F,
12674                                   {X, Y, Builder.CreateFSub(Zero, Z, "sub")});
12675       case PPC::BI__builtin_vsx_xvnmsubadp:
12676       case PPC::BI__builtin_vsx_xvnmsubasp:
12677         Value *FsubRes =
12678           Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")});
12679         return Builder.CreateFSub(Zero, FsubRes, "sub");
12680     }
12681     llvm_unreachable("Unknown FMA operation");
12682     return nullptr; // Suppress no-return warning
12683   }
12684 
12685   case PPC::BI__builtin_vsx_insertword: {
12686     llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxinsertw);
12687 
12688     // Third argument is a compile time constant int. It must be clamped to
12689     // to the range [0, 12].
12690     ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
12691     assert(ArgCI &&
12692            "Third arg to xxinsertw intrinsic must be constant integer");
12693     const int64_t MaxIndex = 12;
12694     int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex);
12695 
12696     // The builtin semantics don't exactly match the xxinsertw instructions
12697     // semantics (which ppc_vsx_xxinsertw follows). The builtin extracts the
12698     // word from the first argument, and inserts it in the second argument. The
12699     // instruction extracts the word from its second input register and inserts
12700     // it into its first input register, so swap the first and second arguments.
12701     std::swap(Ops[0], Ops[1]);
12702 
12703     // Need to cast the second argument from a vector of unsigned int to a
12704     // vector of long long.
12705     Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int64Ty, 2));
12706 
12707     if (getTarget().isLittleEndian()) {
12708       // Create a shuffle mask of (1, 0)
12709       Constant *ShuffleElts[2] = { ConstantInt::get(Int32Ty, 1),
12710                                    ConstantInt::get(Int32Ty, 0)
12711                                  };
12712       Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts);
12713 
12714       // Reverse the double words in the vector we will extract from.
12715       Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2));
12716       Ops[0] = Builder.CreateShuffleVector(Ops[0], Ops[0], ShuffleMask);
12717 
12718       // Reverse the index.
12719       Index = MaxIndex - Index;
12720     }
12721 
12722     // Intrinsic expects the first arg to be a vector of int.
12723     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4));
12724     Ops[2] = ConstantInt::getSigned(Int32Ty, Index);
12725     return Builder.CreateCall(F, Ops);
12726   }
12727 
12728   case PPC::BI__builtin_vsx_extractuword: {
12729     llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxextractuw);
12730 
12731     // Intrinsic expects the first argument to be a vector of doublewords.
12732     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2));
12733 
12734     // The second argument is a compile time constant int that needs to
12735     // be clamped to the range [0, 12].
12736     ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[1]);
12737     assert(ArgCI &&
12738            "Second Arg to xxextractuw intrinsic must be a constant integer!");
12739     const int64_t MaxIndex = 12;
12740     int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex);
12741 
12742     if (getTarget().isLittleEndian()) {
12743       // Reverse the index.
12744       Index = MaxIndex - Index;
12745       Ops[1] = ConstantInt::getSigned(Int32Ty, Index);
12746 
12747       // Emit the call, then reverse the double words of the results vector.
12748       Value *Call = Builder.CreateCall(F, Ops);
12749 
12750       // Create a shuffle mask of (1, 0)
12751       Constant *ShuffleElts[2] = { ConstantInt::get(Int32Ty, 1),
12752                                    ConstantInt::get(Int32Ty, 0)
12753                                  };
12754       Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts);
12755 
12756       Value *ShuffleCall = Builder.CreateShuffleVector(Call, Call, ShuffleMask);
12757       return ShuffleCall;
12758     } else {
12759       Ops[1] = ConstantInt::getSigned(Int32Ty, Index);
12760       return Builder.CreateCall(F, Ops);
12761     }
12762   }
12763 
12764   case PPC::BI__builtin_vsx_xxpermdi: {
12765     ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
12766     assert(ArgCI && "Third arg must be constant integer!");
12767 
12768     unsigned Index = ArgCI->getZExtValue();
12769     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2));
12770     Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int64Ty, 2));
12771 
12772     // Account for endianness by treating this as just a shuffle. So we use the
12773     // same indices for both LE and BE in order to produce expected results in
12774     // both cases.
12775     unsigned ElemIdx0 = (Index & 2) >> 1;
12776     unsigned ElemIdx1 = 2 + (Index & 1);
12777 
12778     Constant *ShuffleElts[2] = {ConstantInt::get(Int32Ty, ElemIdx0),
12779                                 ConstantInt::get(Int32Ty, ElemIdx1)};
12780     Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts);
12781 
12782     Value *ShuffleCall =
12783         Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleMask);
12784     QualType BIRetType = E->getType();
12785     auto RetTy = ConvertType(BIRetType);
12786     return Builder.CreateBitCast(ShuffleCall, RetTy);
12787   }
12788 
12789   case PPC::BI__builtin_vsx_xxsldwi: {
12790     ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
12791     assert(ArgCI && "Third argument must be a compile time constant");
12792     unsigned Index = ArgCI->getZExtValue() & 0x3;
12793     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4));
12794     Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int32Ty, 4));
12795 
12796     // Create a shuffle mask
12797     unsigned ElemIdx0;
12798     unsigned ElemIdx1;
12799     unsigned ElemIdx2;
12800     unsigned ElemIdx3;
12801     if (getTarget().isLittleEndian()) {
12802       // Little endian element N comes from element 8+N-Index of the
12803       // concatenated wide vector (of course, using modulo arithmetic on
12804       // the total number of elements).
12805       ElemIdx0 = (8 - Index) % 8;
12806       ElemIdx1 = (9 - Index) % 8;
12807       ElemIdx2 = (10 - Index) % 8;
12808       ElemIdx3 = (11 - Index) % 8;
12809     } else {
12810       // Big endian ElemIdx<N> = Index + N
12811       ElemIdx0 = Index;
12812       ElemIdx1 = Index + 1;
12813       ElemIdx2 = Index + 2;
12814       ElemIdx3 = Index + 3;
12815     }
12816 
12817     Constant *ShuffleElts[4] = {ConstantInt::get(Int32Ty, ElemIdx0),
12818                                 ConstantInt::get(Int32Ty, ElemIdx1),
12819                                 ConstantInt::get(Int32Ty, ElemIdx2),
12820                                 ConstantInt::get(Int32Ty, ElemIdx3)};
12821 
12822     Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts);
12823     Value *ShuffleCall =
12824         Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleMask);
12825     QualType BIRetType = E->getType();
12826     auto RetTy = ConvertType(BIRetType);
12827     return Builder.CreateBitCast(ShuffleCall, RetTy);
12828   }
12829 
12830   case PPC::BI__builtin_pack_vector_int128: {
12831     bool isLittleEndian = getTarget().isLittleEndian();
12832     Value *UndefValue =
12833         llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), 2));
12834     Value *Res = Builder.CreateInsertElement(
12835         UndefValue, Ops[0], (uint64_t)(isLittleEndian ? 1 : 0));
12836     Res = Builder.CreateInsertElement(Res, Ops[1],
12837                                       (uint64_t)(isLittleEndian ? 0 : 1));
12838     return Builder.CreateBitCast(Res, ConvertType(E->getType()));
12839   }
12840 
12841   case PPC::BI__builtin_unpack_vector_int128: {
12842     ConstantInt *Index = cast<ConstantInt>(Ops[1]);
12843     Value *Unpacked = Builder.CreateBitCast(
12844         Ops[0], llvm::VectorType::get(ConvertType(E->getType()), 2));
12845 
12846     if (getTarget().isLittleEndian())
12847       Index = ConstantInt::get(Index->getType(), 1 - Index->getZExtValue());
12848 
12849     return Builder.CreateExtractElement(Unpacked, Index);
12850   }
12851   }
12852 }
12853 
12854 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
12855                                               const CallExpr *E) {
12856   switch (BuiltinID) {
12857   case AMDGPU::BI__builtin_amdgcn_div_scale:
12858   case AMDGPU::BI__builtin_amdgcn_div_scalef: {
12859     // Translate from the intrinsics's struct return to the builtin's out
12860     // argument.
12861 
12862     Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3));
12863 
12864     llvm::Value *X = EmitScalarExpr(E->getArg(0));
12865     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
12866     llvm::Value *Z = EmitScalarExpr(E->getArg(2));
12867 
12868     llvm::Function *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale,
12869                                            X->getType());
12870 
12871     llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z});
12872 
12873     llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0);
12874     llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1);
12875 
12876     llvm::Type *RealFlagType
12877       = FlagOutPtr.getPointer()->getType()->getPointerElementType();
12878 
12879     llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType);
12880     Builder.CreateStore(FlagExt, FlagOutPtr);
12881     return Result;
12882   }
12883   case AMDGPU::BI__builtin_amdgcn_div_fmas:
12884   case AMDGPU::BI__builtin_amdgcn_div_fmasf: {
12885     llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
12886     llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
12887     llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
12888     llvm::Value *Src3 = EmitScalarExpr(E->getArg(3));
12889 
12890     llvm::Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas,
12891                                       Src0->getType());
12892     llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3);
12893     return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool});
12894   }
12895 
12896   case AMDGPU::BI__builtin_amdgcn_ds_swizzle:
12897     return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_ds_swizzle);
12898   case AMDGPU::BI__builtin_amdgcn_mov_dpp8:
12899     return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_mov_dpp8);
12900   case AMDGPU::BI__builtin_amdgcn_mov_dpp:
12901   case AMDGPU::BI__builtin_amdgcn_update_dpp: {
12902     llvm::SmallVector<llvm::Value *, 6> Args;
12903     for (unsigned I = 0; I != E->getNumArgs(); ++I)
12904       Args.push_back(EmitScalarExpr(E->getArg(I)));
12905     assert(Args.size() == 5 || Args.size() == 6);
12906     if (Args.size() == 5)
12907       Args.insert(Args.begin(), llvm::UndefValue::get(Args[0]->getType()));
12908     Function *F =
12909         CGM.getIntrinsic(Intrinsic::amdgcn_update_dpp, Args[0]->getType());
12910     return Builder.CreateCall(F, Args);
12911   }
12912   case AMDGPU::BI__builtin_amdgcn_div_fixup:
12913   case AMDGPU::BI__builtin_amdgcn_div_fixupf:
12914   case AMDGPU::BI__builtin_amdgcn_div_fixuph:
12915     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_div_fixup);
12916   case AMDGPU::BI__builtin_amdgcn_trig_preop:
12917   case AMDGPU::BI__builtin_amdgcn_trig_preopf:
12918     return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop);
12919   case AMDGPU::BI__builtin_amdgcn_rcp:
12920   case AMDGPU::BI__builtin_amdgcn_rcpf:
12921   case AMDGPU::BI__builtin_amdgcn_rcph:
12922     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp);
12923   case AMDGPU::BI__builtin_amdgcn_rsq:
12924   case AMDGPU::BI__builtin_amdgcn_rsqf:
12925   case AMDGPU::BI__builtin_amdgcn_rsqh:
12926     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq);
12927   case AMDGPU::BI__builtin_amdgcn_rsq_clamp:
12928   case AMDGPU::BI__builtin_amdgcn_rsq_clampf:
12929     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp);
12930   case AMDGPU::BI__builtin_amdgcn_sinf:
12931   case AMDGPU::BI__builtin_amdgcn_sinh:
12932     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin);
12933   case AMDGPU::BI__builtin_amdgcn_cosf:
12934   case AMDGPU::BI__builtin_amdgcn_cosh:
12935     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos);
12936   case AMDGPU::BI__builtin_amdgcn_log_clampf:
12937     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp);
12938   case AMDGPU::BI__builtin_amdgcn_ldexp:
12939   case AMDGPU::BI__builtin_amdgcn_ldexpf:
12940   case AMDGPU::BI__builtin_amdgcn_ldexph:
12941     return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp);
12942   case AMDGPU::BI__builtin_amdgcn_frexp_mant:
12943   case AMDGPU::BI__builtin_amdgcn_frexp_mantf:
12944   case AMDGPU::BI__builtin_amdgcn_frexp_manth:
12945     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant);
12946   case AMDGPU::BI__builtin_amdgcn_frexp_exp:
12947   case AMDGPU::BI__builtin_amdgcn_frexp_expf: {
12948     Value *Src0 = EmitScalarExpr(E->getArg(0));
12949     Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp,
12950                                 { Builder.getInt32Ty(), Src0->getType() });
12951     return Builder.CreateCall(F, Src0);
12952   }
12953   case AMDGPU::BI__builtin_amdgcn_frexp_exph: {
12954     Value *Src0 = EmitScalarExpr(E->getArg(0));
12955     Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp,
12956                                 { Builder.getInt16Ty(), Src0->getType() });
12957     return Builder.CreateCall(F, Src0);
12958   }
12959   case AMDGPU::BI__builtin_amdgcn_fract:
12960   case AMDGPU::BI__builtin_amdgcn_fractf:
12961   case AMDGPU::BI__builtin_amdgcn_fracth:
12962     return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_fract);
12963   case AMDGPU::BI__builtin_amdgcn_lerp:
12964     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_lerp);
12965   case AMDGPU::BI__builtin_amdgcn_ubfe:
12966     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_ubfe);
12967   case AMDGPU::BI__builtin_amdgcn_sbfe:
12968     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_sbfe);
12969   case AMDGPU::BI__builtin_amdgcn_uicmp:
12970   case AMDGPU::BI__builtin_amdgcn_uicmpl:
12971   case AMDGPU::BI__builtin_amdgcn_sicmp:
12972   case AMDGPU::BI__builtin_amdgcn_sicmpl: {
12973     llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
12974     llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
12975     llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
12976 
12977     // FIXME-GFX10: How should 32 bit mask be handled?
12978     Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_icmp,
12979       { Builder.getInt64Ty(), Src0->getType() });
12980     return Builder.CreateCall(F, { Src0, Src1, Src2 });
12981   }
12982   case AMDGPU::BI__builtin_amdgcn_fcmp:
12983   case AMDGPU::BI__builtin_amdgcn_fcmpf: {
12984     llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
12985     llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
12986     llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
12987 
12988     // FIXME-GFX10: How should 32 bit mask be handled?
12989     Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_fcmp,
12990       { Builder.getInt64Ty(), Src0->getType() });
12991     return Builder.CreateCall(F, { Src0, Src1, Src2 });
12992   }
12993   case AMDGPU::BI__builtin_amdgcn_class:
12994   case AMDGPU::BI__builtin_amdgcn_classf:
12995   case AMDGPU::BI__builtin_amdgcn_classh:
12996     return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class);
12997   case AMDGPU::BI__builtin_amdgcn_fmed3f:
12998   case AMDGPU::BI__builtin_amdgcn_fmed3h:
12999     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3);
13000   case AMDGPU::BI__builtin_amdgcn_ds_append:
13001   case AMDGPU::BI__builtin_amdgcn_ds_consume: {
13002     Intrinsic::ID Intrin = BuiltinID == AMDGPU::BI__builtin_amdgcn_ds_append ?
13003       Intrinsic::amdgcn_ds_append : Intrinsic::amdgcn_ds_consume;
13004     Value *Src0 = EmitScalarExpr(E->getArg(0));
13005     Function *F = CGM.getIntrinsic(Intrin, { Src0->getType() });
13006     return Builder.CreateCall(F, { Src0, Builder.getFalse() });
13007   }
13008   case AMDGPU::BI__builtin_amdgcn_read_exec: {
13009     CallInst *CI = cast<CallInst>(
13010       EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, true, "exec"));
13011     CI->setConvergent();
13012     return CI;
13013   }
13014   case AMDGPU::BI__builtin_amdgcn_read_exec_lo:
13015   case AMDGPU::BI__builtin_amdgcn_read_exec_hi: {
13016     StringRef RegName = BuiltinID == AMDGPU::BI__builtin_amdgcn_read_exec_lo ?
13017       "exec_lo" : "exec_hi";
13018     CallInst *CI = cast<CallInst>(
13019       EmitSpecialRegisterBuiltin(*this, E, Int32Ty, Int32Ty, true, RegName));
13020     CI->setConvergent();
13021     return CI;
13022   }
13023   // amdgcn workitem
13024   case AMDGPU::BI__builtin_amdgcn_workitem_id_x:
13025     return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_x, 0, 1024);
13026   case AMDGPU::BI__builtin_amdgcn_workitem_id_y:
13027     return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_y, 0, 1024);
13028   case AMDGPU::BI__builtin_amdgcn_workitem_id_z:
13029     return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_z, 0, 1024);
13030 
13031   // r600 intrinsics
13032   case AMDGPU::BI__builtin_r600_recipsqrt_ieee:
13033   case AMDGPU::BI__builtin_r600_recipsqrt_ieeef:
13034     return emitUnaryBuiltin(*this, E, Intrinsic::r600_recipsqrt_ieee);
13035   case AMDGPU::BI__builtin_r600_read_tidig_x:
13036     return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_x, 0, 1024);
13037   case AMDGPU::BI__builtin_r600_read_tidig_y:
13038     return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_y, 0, 1024);
13039   case AMDGPU::BI__builtin_r600_read_tidig_z:
13040     return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_z, 0, 1024);
13041   default:
13042     return nullptr;
13043   }
13044 }
13045 
13046 /// Handle a SystemZ function in which the final argument is a pointer
13047 /// to an int that receives the post-instruction CC value.  At the LLVM level
13048 /// this is represented as a function that returns a {result, cc} pair.
13049 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF,
13050                                          unsigned IntrinsicID,
13051                                          const CallExpr *E) {
13052   unsigned NumArgs = E->getNumArgs() - 1;
13053   SmallVector<Value *, 8> Args(NumArgs);
13054   for (unsigned I = 0; I < NumArgs; ++I)
13055     Args[I] = CGF.EmitScalarExpr(E->getArg(I));
13056   Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs));
13057   Function *F = CGF.CGM.getIntrinsic(IntrinsicID);
13058   Value *Call = CGF.Builder.CreateCall(F, Args);
13059   Value *CC = CGF.Builder.CreateExtractValue(Call, 1);
13060   CGF.Builder.CreateStore(CC, CCPtr);
13061   return CGF.Builder.CreateExtractValue(Call, 0);
13062 }
13063 
13064 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID,
13065                                                const CallExpr *E) {
13066   switch (BuiltinID) {
13067   case SystemZ::BI__builtin_tbegin: {
13068     Value *TDB = EmitScalarExpr(E->getArg(0));
13069     Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
13070     Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin);
13071     return Builder.CreateCall(F, {TDB, Control});
13072   }
13073   case SystemZ::BI__builtin_tbegin_nofloat: {
13074     Value *TDB = EmitScalarExpr(E->getArg(0));
13075     Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
13076     Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat);
13077     return Builder.CreateCall(F, {TDB, Control});
13078   }
13079   case SystemZ::BI__builtin_tbeginc: {
13080     Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy);
13081     Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08);
13082     Function *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc);
13083     return Builder.CreateCall(F, {TDB, Control});
13084   }
13085   case SystemZ::BI__builtin_tabort: {
13086     Value *Data = EmitScalarExpr(E->getArg(0));
13087     Function *F = CGM.getIntrinsic(Intrinsic::s390_tabort);
13088     return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort"));
13089   }
13090   case SystemZ::BI__builtin_non_tx_store: {
13091     Value *Address = EmitScalarExpr(E->getArg(0));
13092     Value *Data = EmitScalarExpr(E->getArg(1));
13093     Function *F = CGM.getIntrinsic(Intrinsic::s390_ntstg);
13094     return Builder.CreateCall(F, {Data, Address});
13095   }
13096 
13097   // Vector builtins.  Note that most vector builtins are mapped automatically
13098   // to target-specific LLVM intrinsics.  The ones handled specially here can
13099   // be represented via standard LLVM IR, which is preferable to enable common
13100   // LLVM optimizations.
13101 
13102   case SystemZ::BI__builtin_s390_vpopctb:
13103   case SystemZ::BI__builtin_s390_vpopcth:
13104   case SystemZ::BI__builtin_s390_vpopctf:
13105   case SystemZ::BI__builtin_s390_vpopctg: {
13106     llvm::Type *ResultType = ConvertType(E->getType());
13107     Value *X = EmitScalarExpr(E->getArg(0));
13108     Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
13109     return Builder.CreateCall(F, X);
13110   }
13111 
13112   case SystemZ::BI__builtin_s390_vclzb:
13113   case SystemZ::BI__builtin_s390_vclzh:
13114   case SystemZ::BI__builtin_s390_vclzf:
13115   case SystemZ::BI__builtin_s390_vclzg: {
13116     llvm::Type *ResultType = ConvertType(E->getType());
13117     Value *X = EmitScalarExpr(E->getArg(0));
13118     Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
13119     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType);
13120     return Builder.CreateCall(F, {X, Undef});
13121   }
13122 
13123   case SystemZ::BI__builtin_s390_vctzb:
13124   case SystemZ::BI__builtin_s390_vctzh:
13125   case SystemZ::BI__builtin_s390_vctzf:
13126   case SystemZ::BI__builtin_s390_vctzg: {
13127     llvm::Type *ResultType = ConvertType(E->getType());
13128     Value *X = EmitScalarExpr(E->getArg(0));
13129     Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
13130     Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType);
13131     return Builder.CreateCall(F, {X, Undef});
13132   }
13133 
13134   case SystemZ::BI__builtin_s390_vfsqsb:
13135   case SystemZ::BI__builtin_s390_vfsqdb: {
13136     llvm::Type *ResultType = ConvertType(E->getType());
13137     Value *X = EmitScalarExpr(E->getArg(0));
13138     Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
13139     return Builder.CreateCall(F, X);
13140   }
13141   case SystemZ::BI__builtin_s390_vfmasb:
13142   case SystemZ::BI__builtin_s390_vfmadb: {
13143     llvm::Type *ResultType = ConvertType(E->getType());
13144     Value *X = EmitScalarExpr(E->getArg(0));
13145     Value *Y = EmitScalarExpr(E->getArg(1));
13146     Value *Z = EmitScalarExpr(E->getArg(2));
13147     Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
13148     return Builder.CreateCall(F, {X, Y, Z});
13149   }
13150   case SystemZ::BI__builtin_s390_vfmssb:
13151   case SystemZ::BI__builtin_s390_vfmsdb: {
13152     llvm::Type *ResultType = ConvertType(E->getType());
13153     Value *X = EmitScalarExpr(E->getArg(0));
13154     Value *Y = EmitScalarExpr(E->getArg(1));
13155     Value *Z = EmitScalarExpr(E->getArg(2));
13156     Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType);
13157     Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
13158     return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")});
13159   }
13160   case SystemZ::BI__builtin_s390_vfnmasb:
13161   case SystemZ::BI__builtin_s390_vfnmadb: {
13162     llvm::Type *ResultType = ConvertType(E->getType());
13163     Value *X = EmitScalarExpr(E->getArg(0));
13164     Value *Y = EmitScalarExpr(E->getArg(1));
13165     Value *Z = EmitScalarExpr(E->getArg(2));
13166     Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType);
13167     Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
13168     return Builder.CreateFSub(Zero, Builder.CreateCall(F, {X, Y, Z}), "sub");
13169   }
13170   case SystemZ::BI__builtin_s390_vfnmssb:
13171   case SystemZ::BI__builtin_s390_vfnmsdb: {
13172     llvm::Type *ResultType = ConvertType(E->getType());
13173     Value *X = EmitScalarExpr(E->getArg(0));
13174     Value *Y = EmitScalarExpr(E->getArg(1));
13175     Value *Z = EmitScalarExpr(E->getArg(2));
13176     Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType);
13177     Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
13178     Value *NegZ = Builder.CreateFSub(Zero, Z, "sub");
13179     return Builder.CreateFSub(Zero, Builder.CreateCall(F, {X, Y, NegZ}));
13180   }
13181   case SystemZ::BI__builtin_s390_vflpsb:
13182   case SystemZ::BI__builtin_s390_vflpdb: {
13183     llvm::Type *ResultType = ConvertType(E->getType());
13184     Value *X = EmitScalarExpr(E->getArg(0));
13185     Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
13186     return Builder.CreateCall(F, X);
13187   }
13188   case SystemZ::BI__builtin_s390_vflnsb:
13189   case SystemZ::BI__builtin_s390_vflndb: {
13190     llvm::Type *ResultType = ConvertType(E->getType());
13191     Value *X = EmitScalarExpr(E->getArg(0));
13192     Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType);
13193     Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
13194     return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub");
13195   }
13196   case SystemZ::BI__builtin_s390_vfisb:
13197   case SystemZ::BI__builtin_s390_vfidb: {
13198     llvm::Type *ResultType = ConvertType(E->getType());
13199     Value *X = EmitScalarExpr(E->getArg(0));
13200     // Constant-fold the M4 and M5 mask arguments.
13201     llvm::APSInt M4, M5;
13202     bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext());
13203     bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext());
13204     assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?");
13205     (void)IsConstM4; (void)IsConstM5;
13206     // Check whether this instance can be represented via a LLVM standard
13207     // intrinsic.  We only support some combinations of M4 and M5.
13208     Intrinsic::ID ID = Intrinsic::not_intrinsic;
13209     switch (M4.getZExtValue()) {
13210     default: break;
13211     case 0:  // IEEE-inexact exception allowed
13212       switch (M5.getZExtValue()) {
13213       default: break;
13214       case 0: ID = Intrinsic::rint; break;
13215       }
13216       break;
13217     case 4:  // IEEE-inexact exception suppressed
13218       switch (M5.getZExtValue()) {
13219       default: break;
13220       case 0: ID = Intrinsic::nearbyint; break;
13221       case 1: ID = Intrinsic::round; break;
13222       case 5: ID = Intrinsic::trunc; break;
13223       case 6: ID = Intrinsic::ceil; break;
13224       case 7: ID = Intrinsic::floor; break;
13225       }
13226       break;
13227     }
13228     if (ID != Intrinsic::not_intrinsic) {
13229       Function *F = CGM.getIntrinsic(ID, ResultType);
13230       return Builder.CreateCall(F, X);
13231     }
13232     switch (BuiltinID) {
13233       case SystemZ::BI__builtin_s390_vfisb: ID = Intrinsic::s390_vfisb; break;
13234       case SystemZ::BI__builtin_s390_vfidb: ID = Intrinsic::s390_vfidb; break;
13235       default: llvm_unreachable("Unknown BuiltinID");
13236     }
13237     Function *F = CGM.getIntrinsic(ID);
13238     Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
13239     Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5);
13240     return Builder.CreateCall(F, {X, M4Value, M5Value});
13241   }
13242   case SystemZ::BI__builtin_s390_vfmaxsb:
13243   case SystemZ::BI__builtin_s390_vfmaxdb: {
13244     llvm::Type *ResultType = ConvertType(E->getType());
13245     Value *X = EmitScalarExpr(E->getArg(0));
13246     Value *Y = EmitScalarExpr(E->getArg(1));
13247     // Constant-fold the M4 mask argument.
13248     llvm::APSInt M4;
13249     bool IsConstM4 = E->getArg(2)->isIntegerConstantExpr(M4, getContext());
13250     assert(IsConstM4 && "Constant arg isn't actually constant?");
13251     (void)IsConstM4;
13252     // Check whether this instance can be represented via a LLVM standard
13253     // intrinsic.  We only support some values of M4.
13254     Intrinsic::ID ID = Intrinsic::not_intrinsic;
13255     switch (M4.getZExtValue()) {
13256     default: break;
13257     case 4: ID = Intrinsic::maxnum; break;
13258     }
13259     if (ID != Intrinsic::not_intrinsic) {
13260       Function *F = CGM.getIntrinsic(ID, ResultType);
13261       return Builder.CreateCall(F, {X, Y});
13262     }
13263     switch (BuiltinID) {
13264       case SystemZ::BI__builtin_s390_vfmaxsb: ID = Intrinsic::s390_vfmaxsb; break;
13265       case SystemZ::BI__builtin_s390_vfmaxdb: ID = Intrinsic::s390_vfmaxdb; break;
13266       default: llvm_unreachable("Unknown BuiltinID");
13267     }
13268     Function *F = CGM.getIntrinsic(ID);
13269     Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
13270     return Builder.CreateCall(F, {X, Y, M4Value});
13271   }
13272   case SystemZ::BI__builtin_s390_vfminsb:
13273   case SystemZ::BI__builtin_s390_vfmindb: {
13274     llvm::Type *ResultType = ConvertType(E->getType());
13275     Value *X = EmitScalarExpr(E->getArg(0));
13276     Value *Y = EmitScalarExpr(E->getArg(1));
13277     // Constant-fold the M4 mask argument.
13278     llvm::APSInt M4;
13279     bool IsConstM4 = E->getArg(2)->isIntegerConstantExpr(M4, getContext());
13280     assert(IsConstM4 && "Constant arg isn't actually constant?");
13281     (void)IsConstM4;
13282     // Check whether this instance can be represented via a LLVM standard
13283     // intrinsic.  We only support some values of M4.
13284     Intrinsic::ID ID = Intrinsic::not_intrinsic;
13285     switch (M4.getZExtValue()) {
13286     default: break;
13287     case 4: ID = Intrinsic::minnum; break;
13288     }
13289     if (ID != Intrinsic::not_intrinsic) {
13290       Function *F = CGM.getIntrinsic(ID, ResultType);
13291       return Builder.CreateCall(F, {X, Y});
13292     }
13293     switch (BuiltinID) {
13294       case SystemZ::BI__builtin_s390_vfminsb: ID = Intrinsic::s390_vfminsb; break;
13295       case SystemZ::BI__builtin_s390_vfmindb: ID = Intrinsic::s390_vfmindb; break;
13296       default: llvm_unreachable("Unknown BuiltinID");
13297     }
13298     Function *F = CGM.getIntrinsic(ID);
13299     Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
13300     return Builder.CreateCall(F, {X, Y, M4Value});
13301   }
13302 
13303   case SystemZ::BI__builtin_s390_vlbrh:
13304   case SystemZ::BI__builtin_s390_vlbrf:
13305   case SystemZ::BI__builtin_s390_vlbrg: {
13306     llvm::Type *ResultType = ConvertType(E->getType());
13307     Value *X = EmitScalarExpr(E->getArg(0));
13308     Function *F = CGM.getIntrinsic(Intrinsic::bswap, ResultType);
13309     return Builder.CreateCall(F, X);
13310   }
13311 
13312   // Vector intrinsics that output the post-instruction CC value.
13313 
13314 #define INTRINSIC_WITH_CC(NAME) \
13315     case SystemZ::BI__builtin_##NAME: \
13316       return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E)
13317 
13318   INTRINSIC_WITH_CC(s390_vpkshs);
13319   INTRINSIC_WITH_CC(s390_vpksfs);
13320   INTRINSIC_WITH_CC(s390_vpksgs);
13321 
13322   INTRINSIC_WITH_CC(s390_vpklshs);
13323   INTRINSIC_WITH_CC(s390_vpklsfs);
13324   INTRINSIC_WITH_CC(s390_vpklsgs);
13325 
13326   INTRINSIC_WITH_CC(s390_vceqbs);
13327   INTRINSIC_WITH_CC(s390_vceqhs);
13328   INTRINSIC_WITH_CC(s390_vceqfs);
13329   INTRINSIC_WITH_CC(s390_vceqgs);
13330 
13331   INTRINSIC_WITH_CC(s390_vchbs);
13332   INTRINSIC_WITH_CC(s390_vchhs);
13333   INTRINSIC_WITH_CC(s390_vchfs);
13334   INTRINSIC_WITH_CC(s390_vchgs);
13335 
13336   INTRINSIC_WITH_CC(s390_vchlbs);
13337   INTRINSIC_WITH_CC(s390_vchlhs);
13338   INTRINSIC_WITH_CC(s390_vchlfs);
13339   INTRINSIC_WITH_CC(s390_vchlgs);
13340 
13341   INTRINSIC_WITH_CC(s390_vfaebs);
13342   INTRINSIC_WITH_CC(s390_vfaehs);
13343   INTRINSIC_WITH_CC(s390_vfaefs);
13344 
13345   INTRINSIC_WITH_CC(s390_vfaezbs);
13346   INTRINSIC_WITH_CC(s390_vfaezhs);
13347   INTRINSIC_WITH_CC(s390_vfaezfs);
13348 
13349   INTRINSIC_WITH_CC(s390_vfeebs);
13350   INTRINSIC_WITH_CC(s390_vfeehs);
13351   INTRINSIC_WITH_CC(s390_vfeefs);
13352 
13353   INTRINSIC_WITH_CC(s390_vfeezbs);
13354   INTRINSIC_WITH_CC(s390_vfeezhs);
13355   INTRINSIC_WITH_CC(s390_vfeezfs);
13356 
13357   INTRINSIC_WITH_CC(s390_vfenebs);
13358   INTRINSIC_WITH_CC(s390_vfenehs);
13359   INTRINSIC_WITH_CC(s390_vfenefs);
13360 
13361   INTRINSIC_WITH_CC(s390_vfenezbs);
13362   INTRINSIC_WITH_CC(s390_vfenezhs);
13363   INTRINSIC_WITH_CC(s390_vfenezfs);
13364 
13365   INTRINSIC_WITH_CC(s390_vistrbs);
13366   INTRINSIC_WITH_CC(s390_vistrhs);
13367   INTRINSIC_WITH_CC(s390_vistrfs);
13368 
13369   INTRINSIC_WITH_CC(s390_vstrcbs);
13370   INTRINSIC_WITH_CC(s390_vstrchs);
13371   INTRINSIC_WITH_CC(s390_vstrcfs);
13372 
13373   INTRINSIC_WITH_CC(s390_vstrczbs);
13374   INTRINSIC_WITH_CC(s390_vstrczhs);
13375   INTRINSIC_WITH_CC(s390_vstrczfs);
13376 
13377   INTRINSIC_WITH_CC(s390_vfcesbs);
13378   INTRINSIC_WITH_CC(s390_vfcedbs);
13379   INTRINSIC_WITH_CC(s390_vfchsbs);
13380   INTRINSIC_WITH_CC(s390_vfchdbs);
13381   INTRINSIC_WITH_CC(s390_vfchesbs);
13382   INTRINSIC_WITH_CC(s390_vfchedbs);
13383 
13384   INTRINSIC_WITH_CC(s390_vftcisb);
13385   INTRINSIC_WITH_CC(s390_vftcidb);
13386 
13387   INTRINSIC_WITH_CC(s390_vstrsb);
13388   INTRINSIC_WITH_CC(s390_vstrsh);
13389   INTRINSIC_WITH_CC(s390_vstrsf);
13390 
13391   INTRINSIC_WITH_CC(s390_vstrszb);
13392   INTRINSIC_WITH_CC(s390_vstrszh);
13393   INTRINSIC_WITH_CC(s390_vstrszf);
13394 
13395 #undef INTRINSIC_WITH_CC
13396 
13397   default:
13398     return nullptr;
13399   }
13400 }
13401 
13402 namespace {
13403 // Helper classes for mapping MMA builtins to particular LLVM intrinsic variant.
13404 struct NVPTXMmaLdstInfo {
13405   unsigned NumResults;  // Number of elements to load/store
13406   // Intrinsic IDs for row/col variants. 0 if particular layout is unsupported.
13407   unsigned IID_col;
13408   unsigned IID_row;
13409 };
13410 
13411 #define MMA_INTR(geom_op_type, layout) \
13412   Intrinsic::nvvm_wmma_##geom_op_type##_##layout##_stride
13413 #define MMA_LDST(n, geom_op_type)                                              \
13414   { n, MMA_INTR(geom_op_type, col), MMA_INTR(geom_op_type, row) }
13415 
13416 static NVPTXMmaLdstInfo getNVPTXMmaLdstInfo(unsigned BuiltinID) {
13417   switch (BuiltinID) {
13418   // FP MMA loads
13419   case NVPTX::BI__hmma_m16n16k16_ld_a:
13420     return MMA_LDST(8, m16n16k16_load_a_f16);
13421   case NVPTX::BI__hmma_m16n16k16_ld_b:
13422     return MMA_LDST(8, m16n16k16_load_b_f16);
13423   case NVPTX::BI__hmma_m16n16k16_ld_c_f16:
13424     return MMA_LDST(4, m16n16k16_load_c_f16);
13425   case NVPTX::BI__hmma_m16n16k16_ld_c_f32:
13426     return MMA_LDST(8, m16n16k16_load_c_f32);
13427   case NVPTX::BI__hmma_m32n8k16_ld_a:
13428     return MMA_LDST(8, m32n8k16_load_a_f16);
13429   case NVPTX::BI__hmma_m32n8k16_ld_b:
13430     return MMA_LDST(8, m32n8k16_load_b_f16);
13431   case NVPTX::BI__hmma_m32n8k16_ld_c_f16:
13432     return MMA_LDST(4, m32n8k16_load_c_f16);
13433   case NVPTX::BI__hmma_m32n8k16_ld_c_f32:
13434     return MMA_LDST(8, m32n8k16_load_c_f32);
13435   case NVPTX::BI__hmma_m8n32k16_ld_a:
13436     return MMA_LDST(8, m8n32k16_load_a_f16);
13437   case NVPTX::BI__hmma_m8n32k16_ld_b:
13438     return MMA_LDST(8, m8n32k16_load_b_f16);
13439   case NVPTX::BI__hmma_m8n32k16_ld_c_f16:
13440     return MMA_LDST(4, m8n32k16_load_c_f16);
13441   case NVPTX::BI__hmma_m8n32k16_ld_c_f32:
13442     return MMA_LDST(8, m8n32k16_load_c_f32);
13443 
13444   // Integer MMA loads
13445   case NVPTX::BI__imma_m16n16k16_ld_a_s8:
13446     return MMA_LDST(2, m16n16k16_load_a_s8);
13447   case NVPTX::BI__imma_m16n16k16_ld_a_u8:
13448     return MMA_LDST(2, m16n16k16_load_a_u8);
13449   case NVPTX::BI__imma_m16n16k16_ld_b_s8:
13450     return MMA_LDST(2, m16n16k16_load_b_s8);
13451   case NVPTX::BI__imma_m16n16k16_ld_b_u8:
13452     return MMA_LDST(2, m16n16k16_load_b_u8);
13453   case NVPTX::BI__imma_m16n16k16_ld_c:
13454     return MMA_LDST(8, m16n16k16_load_c_s32);
13455   case NVPTX::BI__imma_m32n8k16_ld_a_s8:
13456     return MMA_LDST(4, m32n8k16_load_a_s8);
13457   case NVPTX::BI__imma_m32n8k16_ld_a_u8:
13458     return MMA_LDST(4, m32n8k16_load_a_u8);
13459   case NVPTX::BI__imma_m32n8k16_ld_b_s8:
13460     return MMA_LDST(1, m32n8k16_load_b_s8);
13461   case NVPTX::BI__imma_m32n8k16_ld_b_u8:
13462     return MMA_LDST(1, m32n8k16_load_b_u8);
13463   case NVPTX::BI__imma_m32n8k16_ld_c:
13464     return MMA_LDST(8, m32n8k16_load_c_s32);
13465   case NVPTX::BI__imma_m8n32k16_ld_a_s8:
13466     return MMA_LDST(1, m8n32k16_load_a_s8);
13467   case NVPTX::BI__imma_m8n32k16_ld_a_u8:
13468     return MMA_LDST(1, m8n32k16_load_a_u8);
13469   case NVPTX::BI__imma_m8n32k16_ld_b_s8:
13470     return MMA_LDST(4, m8n32k16_load_b_s8);
13471   case NVPTX::BI__imma_m8n32k16_ld_b_u8:
13472     return MMA_LDST(4, m8n32k16_load_b_u8);
13473   case NVPTX::BI__imma_m8n32k16_ld_c:
13474     return MMA_LDST(8, m8n32k16_load_c_s32);
13475 
13476   // Sub-integer MMA loads.
13477   // Only row/col layout is supported by A/B fragments.
13478   case NVPTX::BI__imma_m8n8k32_ld_a_s4:
13479     return {1, 0, MMA_INTR(m8n8k32_load_a_s4, row)};
13480   case NVPTX::BI__imma_m8n8k32_ld_a_u4:
13481     return {1, 0, MMA_INTR(m8n8k32_load_a_u4, row)};
13482   case NVPTX::BI__imma_m8n8k32_ld_b_s4:
13483     return {1, MMA_INTR(m8n8k32_load_b_s4, col), 0};
13484   case NVPTX::BI__imma_m8n8k32_ld_b_u4:
13485     return {1, MMA_INTR(m8n8k32_load_b_u4, col), 0};
13486   case NVPTX::BI__imma_m8n8k32_ld_c:
13487     return MMA_LDST(2, m8n8k32_load_c_s32);
13488   case NVPTX::BI__bmma_m8n8k128_ld_a_b1:
13489     return {1, 0, MMA_INTR(m8n8k128_load_a_b1, row)};
13490   case NVPTX::BI__bmma_m8n8k128_ld_b_b1:
13491     return {1, MMA_INTR(m8n8k128_load_b_b1, col), 0};
13492   case NVPTX::BI__bmma_m8n8k128_ld_c:
13493     return MMA_LDST(2, m8n8k128_load_c_s32);
13494 
13495   // NOTE: We need to follow inconsitent naming scheme used by NVCC.  Unlike
13496   // PTX and LLVM IR where stores always use fragment D, NVCC builtins always
13497   // use fragment C for both loads and stores.
13498   // FP MMA stores.
13499   case NVPTX::BI__hmma_m16n16k16_st_c_f16:
13500     return MMA_LDST(4, m16n16k16_store_d_f16);
13501   case NVPTX::BI__hmma_m16n16k16_st_c_f32:
13502     return MMA_LDST(8, m16n16k16_store_d_f32);
13503   case NVPTX::BI__hmma_m32n8k16_st_c_f16:
13504     return MMA_LDST(4, m32n8k16_store_d_f16);
13505   case NVPTX::BI__hmma_m32n8k16_st_c_f32:
13506     return MMA_LDST(8, m32n8k16_store_d_f32);
13507   case NVPTX::BI__hmma_m8n32k16_st_c_f16:
13508     return MMA_LDST(4, m8n32k16_store_d_f16);
13509   case NVPTX::BI__hmma_m8n32k16_st_c_f32:
13510     return MMA_LDST(8, m8n32k16_store_d_f32);
13511 
13512   // Integer and sub-integer MMA stores.
13513   // Another naming quirk. Unlike other MMA builtins that use PTX types in the
13514   // name, integer loads/stores use LLVM's i32.
13515   case NVPTX::BI__imma_m16n16k16_st_c_i32:
13516     return MMA_LDST(8, m16n16k16_store_d_s32);
13517   case NVPTX::BI__imma_m32n8k16_st_c_i32:
13518     return MMA_LDST(8, m32n8k16_store_d_s32);
13519   case NVPTX::BI__imma_m8n32k16_st_c_i32:
13520     return MMA_LDST(8, m8n32k16_store_d_s32);
13521   case NVPTX::BI__imma_m8n8k32_st_c_i32:
13522     return MMA_LDST(2, m8n8k32_store_d_s32);
13523   case NVPTX::BI__bmma_m8n8k128_st_c_i32:
13524     return MMA_LDST(2, m8n8k128_store_d_s32);
13525 
13526   default:
13527     llvm_unreachable("Unknown MMA builtin");
13528   }
13529 }
13530 #undef MMA_LDST
13531 #undef MMA_INTR
13532 
13533 
13534 struct NVPTXMmaInfo {
13535   unsigned NumEltsA;
13536   unsigned NumEltsB;
13537   unsigned NumEltsC;
13538   unsigned NumEltsD;
13539   std::array<unsigned, 8> Variants;
13540 
13541   unsigned getMMAIntrinsic(int Layout, bool Satf) {
13542     unsigned Index = Layout * 2 + Satf;
13543     if (Index >= Variants.size())
13544       return 0;
13545     return Variants[Index];
13546   }
13547 };
13548 
13549   // Returns an intrinsic that matches Layout and Satf for valid combinations of
13550   // Layout and Satf, 0 otherwise.
13551 static NVPTXMmaInfo getNVPTXMmaInfo(unsigned BuiltinID) {
13552   // clang-format off
13553 #define MMA_VARIANTS(geom, type) {{                                 \
13554       Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type,             \
13555       Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type##_satfinite, \
13556       Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type,             \
13557       Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \
13558       Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type,             \
13559       Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type##_satfinite, \
13560       Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type,             \
13561       Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type##_satfinite  \
13562     }}
13563 // Sub-integer MMA only supports row.col layout.
13564 #define MMA_VARIANTS_I4(geom, type) {{ \
13565       0, \
13566       0, \
13567       Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type,             \
13568       Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \
13569       0, \
13570       0, \
13571       0, \
13572       0  \
13573     }}
13574 // b1 MMA does not support .satfinite.
13575 #define MMA_VARIANTS_B1(geom, type) {{ \
13576       0, \
13577       0, \
13578       Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type,             \
13579       0, \
13580       0, \
13581       0, \
13582       0, \
13583       0  \
13584     }}
13585     // clang-format on
13586     switch (BuiltinID) {
13587     // FP MMA
13588     // Note that 'type' argument of MMA_VARIANT uses D_C notation, while
13589     // NumEltsN of return value are ordered as A,B,C,D.
13590     case NVPTX::BI__hmma_m16n16k16_mma_f16f16:
13591       return {8, 8, 4, 4, MMA_VARIANTS(m16n16k16, f16_f16)};
13592     case NVPTX::BI__hmma_m16n16k16_mma_f32f16:
13593       return {8, 8, 4, 8, MMA_VARIANTS(m16n16k16, f32_f16)};
13594     case NVPTX::BI__hmma_m16n16k16_mma_f16f32:
13595       return {8, 8, 8, 4, MMA_VARIANTS(m16n16k16, f16_f32)};
13596     case NVPTX::BI__hmma_m16n16k16_mma_f32f32:
13597       return {8, 8, 8, 8, MMA_VARIANTS(m16n16k16, f32_f32)};
13598     case NVPTX::BI__hmma_m32n8k16_mma_f16f16:
13599       return {8, 8, 4, 4, MMA_VARIANTS(m32n8k16, f16_f16)};
13600     case NVPTX::BI__hmma_m32n8k16_mma_f32f16:
13601       return {8, 8, 4, 8, MMA_VARIANTS(m32n8k16, f32_f16)};
13602     case NVPTX::BI__hmma_m32n8k16_mma_f16f32:
13603       return {8, 8, 8, 4, MMA_VARIANTS(m32n8k16, f16_f32)};
13604     case NVPTX::BI__hmma_m32n8k16_mma_f32f32:
13605       return {8, 8, 8, 8, MMA_VARIANTS(m32n8k16, f32_f32)};
13606     case NVPTX::BI__hmma_m8n32k16_mma_f16f16:
13607       return {8, 8, 4, 4, MMA_VARIANTS(m8n32k16, f16_f16)};
13608     case NVPTX::BI__hmma_m8n32k16_mma_f32f16:
13609       return {8, 8, 4, 8, MMA_VARIANTS(m8n32k16, f32_f16)};
13610     case NVPTX::BI__hmma_m8n32k16_mma_f16f32:
13611       return {8, 8, 8, 4, MMA_VARIANTS(m8n32k16, f16_f32)};
13612     case NVPTX::BI__hmma_m8n32k16_mma_f32f32:
13613       return {8, 8, 8, 8, MMA_VARIANTS(m8n32k16, f32_f32)};
13614 
13615     // Integer MMA
13616     case NVPTX::BI__imma_m16n16k16_mma_s8:
13617       return {2, 2, 8, 8, MMA_VARIANTS(m16n16k16, s8)};
13618     case NVPTX::BI__imma_m16n16k16_mma_u8:
13619       return {2, 2, 8, 8, MMA_VARIANTS(m16n16k16, u8)};
13620     case NVPTX::BI__imma_m32n8k16_mma_s8:
13621       return {4, 1, 8, 8, MMA_VARIANTS(m32n8k16, s8)};
13622     case NVPTX::BI__imma_m32n8k16_mma_u8:
13623       return {4, 1, 8, 8, MMA_VARIANTS(m32n8k16, u8)};
13624     case NVPTX::BI__imma_m8n32k16_mma_s8:
13625       return {1, 4, 8, 8, MMA_VARIANTS(m8n32k16, s8)};
13626     case NVPTX::BI__imma_m8n32k16_mma_u8:
13627       return {1, 4, 8, 8, MMA_VARIANTS(m8n32k16, u8)};
13628 
13629     // Sub-integer MMA
13630     case NVPTX::BI__imma_m8n8k32_mma_s4:
13631       return {1, 1, 2, 2, MMA_VARIANTS_I4(m8n8k32, s4)};
13632     case NVPTX::BI__imma_m8n8k32_mma_u4:
13633       return {1, 1, 2, 2, MMA_VARIANTS_I4(m8n8k32, u4)};
13634     case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1:
13635       return {1, 1, 2, 2, MMA_VARIANTS_B1(m8n8k128, b1)};
13636     default:
13637       llvm_unreachable("Unexpected builtin ID.");
13638     }
13639 #undef MMA_VARIANTS
13640 #undef MMA_VARIANTS_I4
13641 #undef MMA_VARIANTS_B1
13642 }
13643 
13644 } // namespace
13645 
13646 Value *
13647 CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, const CallExpr *E) {
13648   auto MakeLdg = [&](unsigned IntrinsicID) {
13649     Value *Ptr = EmitScalarExpr(E->getArg(0));
13650     clang::CharUnits Align =
13651         getNaturalPointeeTypeAlignment(E->getArg(0)->getType());
13652     return Builder.CreateCall(
13653         CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(),
13654                                        Ptr->getType()}),
13655         {Ptr, ConstantInt::get(Builder.getInt32Ty(), Align.getQuantity())});
13656   };
13657   auto MakeScopedAtomic = [&](unsigned IntrinsicID) {
13658     Value *Ptr = EmitScalarExpr(E->getArg(0));
13659     return Builder.CreateCall(
13660         CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(),
13661                                        Ptr->getType()}),
13662         {Ptr, EmitScalarExpr(E->getArg(1))});
13663   };
13664   switch (BuiltinID) {
13665   case NVPTX::BI__nvvm_atom_add_gen_i:
13666   case NVPTX::BI__nvvm_atom_add_gen_l:
13667   case NVPTX::BI__nvvm_atom_add_gen_ll:
13668     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E);
13669 
13670   case NVPTX::BI__nvvm_atom_sub_gen_i:
13671   case NVPTX::BI__nvvm_atom_sub_gen_l:
13672   case NVPTX::BI__nvvm_atom_sub_gen_ll:
13673     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E);
13674 
13675   case NVPTX::BI__nvvm_atom_and_gen_i:
13676   case NVPTX::BI__nvvm_atom_and_gen_l:
13677   case NVPTX::BI__nvvm_atom_and_gen_ll:
13678     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E);
13679 
13680   case NVPTX::BI__nvvm_atom_or_gen_i:
13681   case NVPTX::BI__nvvm_atom_or_gen_l:
13682   case NVPTX::BI__nvvm_atom_or_gen_ll:
13683     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E);
13684 
13685   case NVPTX::BI__nvvm_atom_xor_gen_i:
13686   case NVPTX::BI__nvvm_atom_xor_gen_l:
13687   case NVPTX::BI__nvvm_atom_xor_gen_ll:
13688     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E);
13689 
13690   case NVPTX::BI__nvvm_atom_xchg_gen_i:
13691   case NVPTX::BI__nvvm_atom_xchg_gen_l:
13692   case NVPTX::BI__nvvm_atom_xchg_gen_ll:
13693     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E);
13694 
13695   case NVPTX::BI__nvvm_atom_max_gen_i:
13696   case NVPTX::BI__nvvm_atom_max_gen_l:
13697   case NVPTX::BI__nvvm_atom_max_gen_ll:
13698     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E);
13699 
13700   case NVPTX::BI__nvvm_atom_max_gen_ui:
13701   case NVPTX::BI__nvvm_atom_max_gen_ul:
13702   case NVPTX::BI__nvvm_atom_max_gen_ull:
13703     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E);
13704 
13705   case NVPTX::BI__nvvm_atom_min_gen_i:
13706   case NVPTX::BI__nvvm_atom_min_gen_l:
13707   case NVPTX::BI__nvvm_atom_min_gen_ll:
13708     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E);
13709 
13710   case NVPTX::BI__nvvm_atom_min_gen_ui:
13711   case NVPTX::BI__nvvm_atom_min_gen_ul:
13712   case NVPTX::BI__nvvm_atom_min_gen_ull:
13713     return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E);
13714 
13715   case NVPTX::BI__nvvm_atom_cas_gen_i:
13716   case NVPTX::BI__nvvm_atom_cas_gen_l:
13717   case NVPTX::BI__nvvm_atom_cas_gen_ll:
13718     // __nvvm_atom_cas_gen_* should return the old value rather than the
13719     // success flag.
13720     return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false);
13721 
13722   case NVPTX::BI__nvvm_atom_add_gen_f:
13723   case NVPTX::BI__nvvm_atom_add_gen_d: {
13724     Value *Ptr = EmitScalarExpr(E->getArg(0));
13725     Value *Val = EmitScalarExpr(E->getArg(1));
13726     return Builder.CreateAtomicRMW(llvm::AtomicRMWInst::FAdd, Ptr, Val,
13727                                    AtomicOrdering::SequentiallyConsistent);
13728   }
13729 
13730   case NVPTX::BI__nvvm_atom_inc_gen_ui: {
13731     Value *Ptr = EmitScalarExpr(E->getArg(0));
13732     Value *Val = EmitScalarExpr(E->getArg(1));
13733     Function *FnALI32 =
13734         CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType());
13735     return Builder.CreateCall(FnALI32, {Ptr, Val});
13736   }
13737 
13738   case NVPTX::BI__nvvm_atom_dec_gen_ui: {
13739     Value *Ptr = EmitScalarExpr(E->getArg(0));
13740     Value *Val = EmitScalarExpr(E->getArg(1));
13741     Function *FnALD32 =
13742         CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType());
13743     return Builder.CreateCall(FnALD32, {Ptr, Val});
13744   }
13745 
13746   case NVPTX::BI__nvvm_ldg_c:
13747   case NVPTX::BI__nvvm_ldg_c2:
13748   case NVPTX::BI__nvvm_ldg_c4:
13749   case NVPTX::BI__nvvm_ldg_s:
13750   case NVPTX::BI__nvvm_ldg_s2:
13751   case NVPTX::BI__nvvm_ldg_s4:
13752   case NVPTX::BI__nvvm_ldg_i:
13753   case NVPTX::BI__nvvm_ldg_i2:
13754   case NVPTX::BI__nvvm_ldg_i4:
13755   case NVPTX::BI__nvvm_ldg_l:
13756   case NVPTX::BI__nvvm_ldg_ll:
13757   case NVPTX::BI__nvvm_ldg_ll2:
13758   case NVPTX::BI__nvvm_ldg_uc:
13759   case NVPTX::BI__nvvm_ldg_uc2:
13760   case NVPTX::BI__nvvm_ldg_uc4:
13761   case NVPTX::BI__nvvm_ldg_us:
13762   case NVPTX::BI__nvvm_ldg_us2:
13763   case NVPTX::BI__nvvm_ldg_us4:
13764   case NVPTX::BI__nvvm_ldg_ui:
13765   case NVPTX::BI__nvvm_ldg_ui2:
13766   case NVPTX::BI__nvvm_ldg_ui4:
13767   case NVPTX::BI__nvvm_ldg_ul:
13768   case NVPTX::BI__nvvm_ldg_ull:
13769   case NVPTX::BI__nvvm_ldg_ull2:
13770     // PTX Interoperability section 2.2: "For a vector with an even number of
13771     // elements, its alignment is set to number of elements times the alignment
13772     // of its member: n*alignof(t)."
13773     return MakeLdg(Intrinsic::nvvm_ldg_global_i);
13774   case NVPTX::BI__nvvm_ldg_f:
13775   case NVPTX::BI__nvvm_ldg_f2:
13776   case NVPTX::BI__nvvm_ldg_f4:
13777   case NVPTX::BI__nvvm_ldg_d:
13778   case NVPTX::BI__nvvm_ldg_d2:
13779     return MakeLdg(Intrinsic::nvvm_ldg_global_f);
13780 
13781   case NVPTX::BI__nvvm_atom_cta_add_gen_i:
13782   case NVPTX::BI__nvvm_atom_cta_add_gen_l:
13783   case NVPTX::BI__nvvm_atom_cta_add_gen_ll:
13784     return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_cta);
13785   case NVPTX::BI__nvvm_atom_sys_add_gen_i:
13786   case NVPTX::BI__nvvm_atom_sys_add_gen_l:
13787   case NVPTX::BI__nvvm_atom_sys_add_gen_ll:
13788     return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_sys);
13789   case NVPTX::BI__nvvm_atom_cta_add_gen_f:
13790   case NVPTX::BI__nvvm_atom_cta_add_gen_d:
13791     return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_cta);
13792   case NVPTX::BI__nvvm_atom_sys_add_gen_f:
13793   case NVPTX::BI__nvvm_atom_sys_add_gen_d:
13794     return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_sys);
13795   case NVPTX::BI__nvvm_atom_cta_xchg_gen_i:
13796   case NVPTX::BI__nvvm_atom_cta_xchg_gen_l:
13797   case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll:
13798     return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_cta);
13799   case NVPTX::BI__nvvm_atom_sys_xchg_gen_i:
13800   case NVPTX::BI__nvvm_atom_sys_xchg_gen_l:
13801   case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll:
13802     return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_sys);
13803   case NVPTX::BI__nvvm_atom_cta_max_gen_i:
13804   case NVPTX::BI__nvvm_atom_cta_max_gen_ui:
13805   case NVPTX::BI__nvvm_atom_cta_max_gen_l:
13806   case NVPTX::BI__nvvm_atom_cta_max_gen_ul:
13807   case NVPTX::BI__nvvm_atom_cta_max_gen_ll:
13808   case NVPTX::BI__nvvm_atom_cta_max_gen_ull:
13809     return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_cta);
13810   case NVPTX::BI__nvvm_atom_sys_max_gen_i:
13811   case NVPTX::BI__nvvm_atom_sys_max_gen_ui:
13812   case NVPTX::BI__nvvm_atom_sys_max_gen_l:
13813   case NVPTX::BI__nvvm_atom_sys_max_gen_ul:
13814   case NVPTX::BI__nvvm_atom_sys_max_gen_ll:
13815   case NVPTX::BI__nvvm_atom_sys_max_gen_ull:
13816     return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_sys);
13817   case NVPTX::BI__nvvm_atom_cta_min_gen_i:
13818   case NVPTX::BI__nvvm_atom_cta_min_gen_ui:
13819   case NVPTX::BI__nvvm_atom_cta_min_gen_l:
13820   case NVPTX::BI__nvvm_atom_cta_min_gen_ul:
13821   case NVPTX::BI__nvvm_atom_cta_min_gen_ll:
13822   case NVPTX::BI__nvvm_atom_cta_min_gen_ull:
13823     return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_cta);
13824   case NVPTX::BI__nvvm_atom_sys_min_gen_i:
13825   case NVPTX::BI__nvvm_atom_sys_min_gen_ui:
13826   case NVPTX::BI__nvvm_atom_sys_min_gen_l:
13827   case NVPTX::BI__nvvm_atom_sys_min_gen_ul:
13828   case NVPTX::BI__nvvm_atom_sys_min_gen_ll:
13829   case NVPTX::BI__nvvm_atom_sys_min_gen_ull:
13830     return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_sys);
13831   case NVPTX::BI__nvvm_atom_cta_inc_gen_ui:
13832     return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_cta);
13833   case NVPTX::BI__nvvm_atom_cta_dec_gen_ui:
13834     return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_cta);
13835   case NVPTX::BI__nvvm_atom_sys_inc_gen_ui:
13836     return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_sys);
13837   case NVPTX::BI__nvvm_atom_sys_dec_gen_ui:
13838     return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_sys);
13839   case NVPTX::BI__nvvm_atom_cta_and_gen_i:
13840   case NVPTX::BI__nvvm_atom_cta_and_gen_l:
13841   case NVPTX::BI__nvvm_atom_cta_and_gen_ll:
13842     return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_cta);
13843   case NVPTX::BI__nvvm_atom_sys_and_gen_i:
13844   case NVPTX::BI__nvvm_atom_sys_and_gen_l:
13845   case NVPTX::BI__nvvm_atom_sys_and_gen_ll:
13846     return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_sys);
13847   case NVPTX::BI__nvvm_atom_cta_or_gen_i:
13848   case NVPTX::BI__nvvm_atom_cta_or_gen_l:
13849   case NVPTX::BI__nvvm_atom_cta_or_gen_ll:
13850     return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_cta);
13851   case NVPTX::BI__nvvm_atom_sys_or_gen_i:
13852   case NVPTX::BI__nvvm_atom_sys_or_gen_l:
13853   case NVPTX::BI__nvvm_atom_sys_or_gen_ll:
13854     return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_sys);
13855   case NVPTX::BI__nvvm_atom_cta_xor_gen_i:
13856   case NVPTX::BI__nvvm_atom_cta_xor_gen_l:
13857   case NVPTX::BI__nvvm_atom_cta_xor_gen_ll:
13858     return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_cta);
13859   case NVPTX::BI__nvvm_atom_sys_xor_gen_i:
13860   case NVPTX::BI__nvvm_atom_sys_xor_gen_l:
13861   case NVPTX::BI__nvvm_atom_sys_xor_gen_ll:
13862     return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys);
13863   case NVPTX::BI__nvvm_atom_cta_cas_gen_i:
13864   case NVPTX::BI__nvvm_atom_cta_cas_gen_l:
13865   case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: {
13866     Value *Ptr = EmitScalarExpr(E->getArg(0));
13867     return Builder.CreateCall(
13868         CGM.getIntrinsic(
13869             Intrinsic::nvvm_atomic_cas_gen_i_cta,
13870             {Ptr->getType()->getPointerElementType(), Ptr->getType()}),
13871         {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
13872   }
13873   case NVPTX::BI__nvvm_atom_sys_cas_gen_i:
13874   case NVPTX::BI__nvvm_atom_sys_cas_gen_l:
13875   case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: {
13876     Value *Ptr = EmitScalarExpr(E->getArg(0));
13877     return Builder.CreateCall(
13878         CGM.getIntrinsic(
13879             Intrinsic::nvvm_atomic_cas_gen_i_sys,
13880             {Ptr->getType()->getPointerElementType(), Ptr->getType()}),
13881         {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
13882   }
13883   case NVPTX::BI__nvvm_match_all_sync_i32p:
13884   case NVPTX::BI__nvvm_match_all_sync_i64p: {
13885     Value *Mask = EmitScalarExpr(E->getArg(0));
13886     Value *Val = EmitScalarExpr(E->getArg(1));
13887     Address PredOutPtr = EmitPointerWithAlignment(E->getArg(2));
13888     Value *ResultPair = Builder.CreateCall(
13889         CGM.getIntrinsic(BuiltinID == NVPTX::BI__nvvm_match_all_sync_i32p
13890                              ? Intrinsic::nvvm_match_all_sync_i32p
13891                              : Intrinsic::nvvm_match_all_sync_i64p),
13892         {Mask, Val});
13893     Value *Pred = Builder.CreateZExt(Builder.CreateExtractValue(ResultPair, 1),
13894                                      PredOutPtr.getElementType());
13895     Builder.CreateStore(Pred, PredOutPtr);
13896     return Builder.CreateExtractValue(ResultPair, 0);
13897   }
13898 
13899   // FP MMA loads
13900   case NVPTX::BI__hmma_m16n16k16_ld_a:
13901   case NVPTX::BI__hmma_m16n16k16_ld_b:
13902   case NVPTX::BI__hmma_m16n16k16_ld_c_f16:
13903   case NVPTX::BI__hmma_m16n16k16_ld_c_f32:
13904   case NVPTX::BI__hmma_m32n8k16_ld_a:
13905   case NVPTX::BI__hmma_m32n8k16_ld_b:
13906   case NVPTX::BI__hmma_m32n8k16_ld_c_f16:
13907   case NVPTX::BI__hmma_m32n8k16_ld_c_f32:
13908   case NVPTX::BI__hmma_m8n32k16_ld_a:
13909   case NVPTX::BI__hmma_m8n32k16_ld_b:
13910   case NVPTX::BI__hmma_m8n32k16_ld_c_f16:
13911   case NVPTX::BI__hmma_m8n32k16_ld_c_f32:
13912   // Integer MMA loads.
13913   case NVPTX::BI__imma_m16n16k16_ld_a_s8:
13914   case NVPTX::BI__imma_m16n16k16_ld_a_u8:
13915   case NVPTX::BI__imma_m16n16k16_ld_b_s8:
13916   case NVPTX::BI__imma_m16n16k16_ld_b_u8:
13917   case NVPTX::BI__imma_m16n16k16_ld_c:
13918   case NVPTX::BI__imma_m32n8k16_ld_a_s8:
13919   case NVPTX::BI__imma_m32n8k16_ld_a_u8:
13920   case NVPTX::BI__imma_m32n8k16_ld_b_s8:
13921   case NVPTX::BI__imma_m32n8k16_ld_b_u8:
13922   case NVPTX::BI__imma_m32n8k16_ld_c:
13923   case NVPTX::BI__imma_m8n32k16_ld_a_s8:
13924   case NVPTX::BI__imma_m8n32k16_ld_a_u8:
13925   case NVPTX::BI__imma_m8n32k16_ld_b_s8:
13926   case NVPTX::BI__imma_m8n32k16_ld_b_u8:
13927   case NVPTX::BI__imma_m8n32k16_ld_c:
13928   // Sub-integer MMA loads.
13929   case NVPTX::BI__imma_m8n8k32_ld_a_s4:
13930   case NVPTX::BI__imma_m8n8k32_ld_a_u4:
13931   case NVPTX::BI__imma_m8n8k32_ld_b_s4:
13932   case NVPTX::BI__imma_m8n8k32_ld_b_u4:
13933   case NVPTX::BI__imma_m8n8k32_ld_c:
13934   case NVPTX::BI__bmma_m8n8k128_ld_a_b1:
13935   case NVPTX::BI__bmma_m8n8k128_ld_b_b1:
13936   case NVPTX::BI__bmma_m8n8k128_ld_c:
13937   {
13938     Address Dst = EmitPointerWithAlignment(E->getArg(0));
13939     Value *Src = EmitScalarExpr(E->getArg(1));
13940     Value *Ldm = EmitScalarExpr(E->getArg(2));
13941     llvm::APSInt isColMajorArg;
13942     if (!E->getArg(3)->isIntegerConstantExpr(isColMajorArg, getContext()))
13943       return nullptr;
13944     bool isColMajor = isColMajorArg.getSExtValue();
13945     NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID);
13946     unsigned IID = isColMajor ? II.IID_col : II.IID_row;
13947     if (IID == 0)
13948       return nullptr;
13949 
13950     Value *Result =
13951         Builder.CreateCall(CGM.getIntrinsic(IID, Src->getType()), {Src, Ldm});
13952 
13953     // Save returned values.
13954     assert(II.NumResults);
13955     if (II.NumResults == 1) {
13956       Builder.CreateAlignedStore(Result, Dst.getPointer(),
13957                                  CharUnits::fromQuantity(4));
13958     } else {
13959       for (unsigned i = 0; i < II.NumResults; ++i) {
13960         Builder.CreateAlignedStore(
13961             Builder.CreateBitCast(Builder.CreateExtractValue(Result, i),
13962                                   Dst.getElementType()),
13963             Builder.CreateGEP(Dst.getPointer(),
13964                               llvm::ConstantInt::get(IntTy, i)),
13965             CharUnits::fromQuantity(4));
13966       }
13967     }
13968     return Result;
13969   }
13970 
13971   case NVPTX::BI__hmma_m16n16k16_st_c_f16:
13972   case NVPTX::BI__hmma_m16n16k16_st_c_f32:
13973   case NVPTX::BI__hmma_m32n8k16_st_c_f16:
13974   case NVPTX::BI__hmma_m32n8k16_st_c_f32:
13975   case NVPTX::BI__hmma_m8n32k16_st_c_f16:
13976   case NVPTX::BI__hmma_m8n32k16_st_c_f32:
13977   case NVPTX::BI__imma_m16n16k16_st_c_i32:
13978   case NVPTX::BI__imma_m32n8k16_st_c_i32:
13979   case NVPTX::BI__imma_m8n32k16_st_c_i32:
13980   case NVPTX::BI__imma_m8n8k32_st_c_i32:
13981   case NVPTX::BI__bmma_m8n8k128_st_c_i32: {
13982     Value *Dst = EmitScalarExpr(E->getArg(0));
13983     Address Src = EmitPointerWithAlignment(E->getArg(1));
13984     Value *Ldm = EmitScalarExpr(E->getArg(2));
13985     llvm::APSInt isColMajorArg;
13986     if (!E->getArg(3)->isIntegerConstantExpr(isColMajorArg, getContext()))
13987       return nullptr;
13988     bool isColMajor = isColMajorArg.getSExtValue();
13989     NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID);
13990     unsigned IID = isColMajor ? II.IID_col : II.IID_row;
13991     if (IID == 0)
13992       return nullptr;
13993     Function *Intrinsic =
13994         CGM.getIntrinsic(IID, Dst->getType());
13995     llvm::Type *ParamType = Intrinsic->getFunctionType()->getParamType(1);
13996     SmallVector<Value *, 10> Values = {Dst};
13997     for (unsigned i = 0; i < II.NumResults; ++i) {
13998       Value *V = Builder.CreateAlignedLoad(
13999           Builder.CreateGEP(Src.getPointer(), llvm::ConstantInt::get(IntTy, i)),
14000           CharUnits::fromQuantity(4));
14001       Values.push_back(Builder.CreateBitCast(V, ParamType));
14002     }
14003     Values.push_back(Ldm);
14004     Value *Result = Builder.CreateCall(Intrinsic, Values);
14005     return Result;
14006   }
14007 
14008   // BI__hmma_m16n16k16_mma_<Dtype><CType>(d, a, b, c, layout, satf) -->
14009   // Intrinsic::nvvm_wmma_m16n16k16_mma_sync<layout A,B><DType><CType><Satf>
14010   case NVPTX::BI__hmma_m16n16k16_mma_f16f16:
14011   case NVPTX::BI__hmma_m16n16k16_mma_f32f16:
14012   case NVPTX::BI__hmma_m16n16k16_mma_f32f32:
14013   case NVPTX::BI__hmma_m16n16k16_mma_f16f32:
14014   case NVPTX::BI__hmma_m32n8k16_mma_f16f16:
14015   case NVPTX::BI__hmma_m32n8k16_mma_f32f16:
14016   case NVPTX::BI__hmma_m32n8k16_mma_f32f32:
14017   case NVPTX::BI__hmma_m32n8k16_mma_f16f32:
14018   case NVPTX::BI__hmma_m8n32k16_mma_f16f16:
14019   case NVPTX::BI__hmma_m8n32k16_mma_f32f16:
14020   case NVPTX::BI__hmma_m8n32k16_mma_f32f32:
14021   case NVPTX::BI__hmma_m8n32k16_mma_f16f32:
14022   case NVPTX::BI__imma_m16n16k16_mma_s8:
14023   case NVPTX::BI__imma_m16n16k16_mma_u8:
14024   case NVPTX::BI__imma_m32n8k16_mma_s8:
14025   case NVPTX::BI__imma_m32n8k16_mma_u8:
14026   case NVPTX::BI__imma_m8n32k16_mma_s8:
14027   case NVPTX::BI__imma_m8n32k16_mma_u8:
14028   case NVPTX::BI__imma_m8n8k32_mma_s4:
14029   case NVPTX::BI__imma_m8n8k32_mma_u4:
14030   case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1: {
14031     Address Dst = EmitPointerWithAlignment(E->getArg(0));
14032     Address SrcA = EmitPointerWithAlignment(E->getArg(1));
14033     Address SrcB = EmitPointerWithAlignment(E->getArg(2));
14034     Address SrcC = EmitPointerWithAlignment(E->getArg(3));
14035     llvm::APSInt LayoutArg;
14036     if (!E->getArg(4)->isIntegerConstantExpr(LayoutArg, getContext()))
14037       return nullptr;
14038     int Layout = LayoutArg.getSExtValue();
14039     if (Layout < 0 || Layout > 3)
14040       return nullptr;
14041     llvm::APSInt SatfArg;
14042     if (BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1)
14043       SatfArg = 0;  // .b1 does not have satf argument.
14044     else if (!E->getArg(5)->isIntegerConstantExpr(SatfArg, getContext()))
14045       return nullptr;
14046     bool Satf = SatfArg.getSExtValue();
14047     NVPTXMmaInfo MI = getNVPTXMmaInfo(BuiltinID);
14048     unsigned IID = MI.getMMAIntrinsic(Layout, Satf);
14049     if (IID == 0)  // Unsupported combination of Layout/Satf.
14050       return nullptr;
14051 
14052     SmallVector<Value *, 24> Values;
14053     Function *Intrinsic = CGM.getIntrinsic(IID);
14054     llvm::Type *AType = Intrinsic->getFunctionType()->getParamType(0);
14055     // Load A
14056     for (unsigned i = 0; i < MI.NumEltsA; ++i) {
14057       Value *V = Builder.CreateAlignedLoad(
14058           Builder.CreateGEP(SrcA.getPointer(),
14059                             llvm::ConstantInt::get(IntTy, i)),
14060           CharUnits::fromQuantity(4));
14061       Values.push_back(Builder.CreateBitCast(V, AType));
14062     }
14063     // Load B
14064     llvm::Type *BType = Intrinsic->getFunctionType()->getParamType(MI.NumEltsA);
14065     for (unsigned i = 0; i < MI.NumEltsB; ++i) {
14066       Value *V = Builder.CreateAlignedLoad(
14067           Builder.CreateGEP(SrcB.getPointer(),
14068                             llvm::ConstantInt::get(IntTy, i)),
14069           CharUnits::fromQuantity(4));
14070       Values.push_back(Builder.CreateBitCast(V, BType));
14071     }
14072     // Load C
14073     llvm::Type *CType =
14074         Intrinsic->getFunctionType()->getParamType(MI.NumEltsA + MI.NumEltsB);
14075     for (unsigned i = 0; i < MI.NumEltsC; ++i) {
14076       Value *V = Builder.CreateAlignedLoad(
14077           Builder.CreateGEP(SrcC.getPointer(),
14078                             llvm::ConstantInt::get(IntTy, i)),
14079           CharUnits::fromQuantity(4));
14080       Values.push_back(Builder.CreateBitCast(V, CType));
14081     }
14082     Value *Result = Builder.CreateCall(Intrinsic, Values);
14083     llvm::Type *DType = Dst.getElementType();
14084     for (unsigned i = 0; i < MI.NumEltsD; ++i)
14085       Builder.CreateAlignedStore(
14086           Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), DType),
14087           Builder.CreateGEP(Dst.getPointer(), llvm::ConstantInt::get(IntTy, i)),
14088           CharUnits::fromQuantity(4));
14089     return Result;
14090   }
14091   default:
14092     return nullptr;
14093   }
14094 }
14095 
14096 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
14097                                                    const CallExpr *E) {
14098   switch (BuiltinID) {
14099   case WebAssembly::BI__builtin_wasm_memory_size: {
14100     llvm::Type *ResultType = ConvertType(E->getType());
14101     Value *I = EmitScalarExpr(E->getArg(0));
14102     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType);
14103     return Builder.CreateCall(Callee, I);
14104   }
14105   case WebAssembly::BI__builtin_wasm_memory_grow: {
14106     llvm::Type *ResultType = ConvertType(E->getType());
14107     Value *Args[] = {
14108       EmitScalarExpr(E->getArg(0)),
14109       EmitScalarExpr(E->getArg(1))
14110     };
14111     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_grow, ResultType);
14112     return Builder.CreateCall(Callee, Args);
14113   }
14114   case WebAssembly::BI__builtin_wasm_memory_init: {
14115     llvm::APSInt SegConst;
14116     if (!E->getArg(0)->isIntegerConstantExpr(SegConst, getContext()))
14117       llvm_unreachable("Constant arg isn't actually constant?");
14118     llvm::APSInt MemConst;
14119     if (!E->getArg(1)->isIntegerConstantExpr(MemConst, getContext()))
14120       llvm_unreachable("Constant arg isn't actually constant?");
14121     if (!MemConst.isNullValue())
14122       ErrorUnsupported(E, "non-zero memory index");
14123     Value *Args[] = {llvm::ConstantInt::get(getLLVMContext(), SegConst),
14124                      llvm::ConstantInt::get(getLLVMContext(), MemConst),
14125                      EmitScalarExpr(E->getArg(2)), EmitScalarExpr(E->getArg(3)),
14126                      EmitScalarExpr(E->getArg(4))};
14127     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_init);
14128     return Builder.CreateCall(Callee, Args);
14129   }
14130   case WebAssembly::BI__builtin_wasm_data_drop: {
14131     llvm::APSInt SegConst;
14132     if (!E->getArg(0)->isIntegerConstantExpr(SegConst, getContext()))
14133       llvm_unreachable("Constant arg isn't actually constant?");
14134     Value *Arg = llvm::ConstantInt::get(getLLVMContext(), SegConst);
14135     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_data_drop);
14136     return Builder.CreateCall(Callee, {Arg});
14137   }
14138   case WebAssembly::BI__builtin_wasm_tls_size: {
14139     llvm::Type *ResultType = ConvertType(E->getType());
14140     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_size, ResultType);
14141     return Builder.CreateCall(Callee);
14142   }
14143   case WebAssembly::BI__builtin_wasm_tls_align: {
14144     llvm::Type *ResultType = ConvertType(E->getType());
14145     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_align, ResultType);
14146     return Builder.CreateCall(Callee);
14147   }
14148   case WebAssembly::BI__builtin_wasm_tls_base: {
14149     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_base);
14150     return Builder.CreateCall(Callee);
14151   }
14152   case WebAssembly::BI__builtin_wasm_throw: {
14153     Value *Tag = EmitScalarExpr(E->getArg(0));
14154     Value *Obj = EmitScalarExpr(E->getArg(1));
14155     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_throw);
14156     return Builder.CreateCall(Callee, {Tag, Obj});
14157   }
14158   case WebAssembly::BI__builtin_wasm_rethrow_in_catch: {
14159     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_rethrow_in_catch);
14160     return Builder.CreateCall(Callee);
14161   }
14162   case WebAssembly::BI__builtin_wasm_atomic_wait_i32: {
14163     Value *Addr = EmitScalarExpr(E->getArg(0));
14164     Value *Expected = EmitScalarExpr(E->getArg(1));
14165     Value *Timeout = EmitScalarExpr(E->getArg(2));
14166     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_wait_i32);
14167     return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
14168   }
14169   case WebAssembly::BI__builtin_wasm_atomic_wait_i64: {
14170     Value *Addr = EmitScalarExpr(E->getArg(0));
14171     Value *Expected = EmitScalarExpr(E->getArg(1));
14172     Value *Timeout = EmitScalarExpr(E->getArg(2));
14173     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_wait_i64);
14174     return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
14175   }
14176   case WebAssembly::BI__builtin_wasm_atomic_notify: {
14177     Value *Addr = EmitScalarExpr(E->getArg(0));
14178     Value *Count = EmitScalarExpr(E->getArg(1));
14179     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_notify);
14180     return Builder.CreateCall(Callee, {Addr, Count});
14181   }
14182   case WebAssembly::BI__builtin_wasm_trunc_s_i32_f32:
14183   case WebAssembly::BI__builtin_wasm_trunc_s_i32_f64:
14184   case WebAssembly::BI__builtin_wasm_trunc_s_i64_f32:
14185   case WebAssembly::BI__builtin_wasm_trunc_s_i64_f64: {
14186     Value *Src = EmitScalarExpr(E->getArg(0));
14187     llvm::Type *ResT = ConvertType(E->getType());
14188     Function *Callee =
14189         CGM.getIntrinsic(Intrinsic::wasm_trunc_signed, {ResT, Src->getType()});
14190     return Builder.CreateCall(Callee, {Src});
14191   }
14192   case WebAssembly::BI__builtin_wasm_trunc_u_i32_f32:
14193   case WebAssembly::BI__builtin_wasm_trunc_u_i32_f64:
14194   case WebAssembly::BI__builtin_wasm_trunc_u_i64_f32:
14195   case WebAssembly::BI__builtin_wasm_trunc_u_i64_f64: {
14196     Value *Src = EmitScalarExpr(E->getArg(0));
14197     llvm::Type *ResT = ConvertType(E->getType());
14198     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_unsigned,
14199                                         {ResT, Src->getType()});
14200     return Builder.CreateCall(Callee, {Src});
14201   }
14202   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f32:
14203   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f64:
14204   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f32:
14205   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f64:
14206   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32x4_f32x4:
14207   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64x2_f64x2: {
14208     Value *Src = EmitScalarExpr(E->getArg(0));
14209     llvm::Type *ResT = ConvertType(E->getType());
14210     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_saturate_signed,
14211                                      {ResT, Src->getType()});
14212     return Builder.CreateCall(Callee, {Src});
14213   }
14214   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f32:
14215   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f64:
14216   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f32:
14217   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f64:
14218   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32x4_f32x4:
14219   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64x2_f64x2: {
14220     Value *Src = EmitScalarExpr(E->getArg(0));
14221     llvm::Type *ResT = ConvertType(E->getType());
14222     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_saturate_unsigned,
14223                                      {ResT, Src->getType()});
14224     return Builder.CreateCall(Callee, {Src});
14225   }
14226   case WebAssembly::BI__builtin_wasm_min_f32:
14227   case WebAssembly::BI__builtin_wasm_min_f64:
14228   case WebAssembly::BI__builtin_wasm_min_f32x4:
14229   case WebAssembly::BI__builtin_wasm_min_f64x2: {
14230     Value *LHS = EmitScalarExpr(E->getArg(0));
14231     Value *RHS = EmitScalarExpr(E->getArg(1));
14232     Function *Callee = CGM.getIntrinsic(Intrinsic::minimum,
14233                                      ConvertType(E->getType()));
14234     return Builder.CreateCall(Callee, {LHS, RHS});
14235   }
14236   case WebAssembly::BI__builtin_wasm_max_f32:
14237   case WebAssembly::BI__builtin_wasm_max_f64:
14238   case WebAssembly::BI__builtin_wasm_max_f32x4:
14239   case WebAssembly::BI__builtin_wasm_max_f64x2: {
14240     Value *LHS = EmitScalarExpr(E->getArg(0));
14241     Value *RHS = EmitScalarExpr(E->getArg(1));
14242     Function *Callee = CGM.getIntrinsic(Intrinsic::maximum,
14243                                      ConvertType(E->getType()));
14244     return Builder.CreateCall(Callee, {LHS, RHS});
14245   }
14246   case WebAssembly::BI__builtin_wasm_swizzle_v8x16: {
14247     Value *Src = EmitScalarExpr(E->getArg(0));
14248     Value *Indices = EmitScalarExpr(E->getArg(1));
14249     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_swizzle);
14250     return Builder.CreateCall(Callee, {Src, Indices});
14251   }
14252   case WebAssembly::BI__builtin_wasm_extract_lane_s_i8x16:
14253   case WebAssembly::BI__builtin_wasm_extract_lane_u_i8x16:
14254   case WebAssembly::BI__builtin_wasm_extract_lane_s_i16x8:
14255   case WebAssembly::BI__builtin_wasm_extract_lane_u_i16x8:
14256   case WebAssembly::BI__builtin_wasm_extract_lane_i32x4:
14257   case WebAssembly::BI__builtin_wasm_extract_lane_i64x2:
14258   case WebAssembly::BI__builtin_wasm_extract_lane_f32x4:
14259   case WebAssembly::BI__builtin_wasm_extract_lane_f64x2: {
14260     llvm::APSInt LaneConst;
14261     if (!E->getArg(1)->isIntegerConstantExpr(LaneConst, getContext()))
14262       llvm_unreachable("Constant arg isn't actually constant?");
14263     Value *Vec = EmitScalarExpr(E->getArg(0));
14264     Value *Lane = llvm::ConstantInt::get(getLLVMContext(), LaneConst);
14265     Value *Extract = Builder.CreateExtractElement(Vec, Lane);
14266     switch (BuiltinID) {
14267     case WebAssembly::BI__builtin_wasm_extract_lane_s_i8x16:
14268     case WebAssembly::BI__builtin_wasm_extract_lane_s_i16x8:
14269       return Builder.CreateSExt(Extract, ConvertType(E->getType()));
14270     case WebAssembly::BI__builtin_wasm_extract_lane_u_i8x16:
14271     case WebAssembly::BI__builtin_wasm_extract_lane_u_i16x8:
14272       return Builder.CreateZExt(Extract, ConvertType(E->getType()));
14273     case WebAssembly::BI__builtin_wasm_extract_lane_i32x4:
14274     case WebAssembly::BI__builtin_wasm_extract_lane_i64x2:
14275     case WebAssembly::BI__builtin_wasm_extract_lane_f32x4:
14276     case WebAssembly::BI__builtin_wasm_extract_lane_f64x2:
14277       return Extract;
14278     default:
14279       llvm_unreachable("unexpected builtin ID");
14280     }
14281   }
14282   case WebAssembly::BI__builtin_wasm_replace_lane_i8x16:
14283   case WebAssembly::BI__builtin_wasm_replace_lane_i16x8:
14284   case WebAssembly::BI__builtin_wasm_replace_lane_i32x4:
14285   case WebAssembly::BI__builtin_wasm_replace_lane_i64x2:
14286   case WebAssembly::BI__builtin_wasm_replace_lane_f32x4:
14287   case WebAssembly::BI__builtin_wasm_replace_lane_f64x2: {
14288     llvm::APSInt LaneConst;
14289     if (!E->getArg(1)->isIntegerConstantExpr(LaneConst, getContext()))
14290       llvm_unreachable("Constant arg isn't actually constant?");
14291     Value *Vec = EmitScalarExpr(E->getArg(0));
14292     Value *Lane = llvm::ConstantInt::get(getLLVMContext(), LaneConst);
14293     Value *Val = EmitScalarExpr(E->getArg(2));
14294     switch (BuiltinID) {
14295     case WebAssembly::BI__builtin_wasm_replace_lane_i8x16:
14296     case WebAssembly::BI__builtin_wasm_replace_lane_i16x8: {
14297       llvm::Type *ElemType = ConvertType(E->getType())->getVectorElementType();
14298       Value *Trunc = Builder.CreateTrunc(Val, ElemType);
14299       return Builder.CreateInsertElement(Vec, Trunc, Lane);
14300     }
14301     case WebAssembly::BI__builtin_wasm_replace_lane_i32x4:
14302     case WebAssembly::BI__builtin_wasm_replace_lane_i64x2:
14303     case WebAssembly::BI__builtin_wasm_replace_lane_f32x4:
14304     case WebAssembly::BI__builtin_wasm_replace_lane_f64x2:
14305       return Builder.CreateInsertElement(Vec, Val, Lane);
14306     default:
14307       llvm_unreachable("unexpected builtin ID");
14308     }
14309   }
14310   case WebAssembly::BI__builtin_wasm_add_saturate_s_i8x16:
14311   case WebAssembly::BI__builtin_wasm_add_saturate_u_i8x16:
14312   case WebAssembly::BI__builtin_wasm_add_saturate_s_i16x8:
14313   case WebAssembly::BI__builtin_wasm_add_saturate_u_i16x8:
14314   case WebAssembly::BI__builtin_wasm_sub_saturate_s_i8x16:
14315   case WebAssembly::BI__builtin_wasm_sub_saturate_u_i8x16:
14316   case WebAssembly::BI__builtin_wasm_sub_saturate_s_i16x8:
14317   case WebAssembly::BI__builtin_wasm_sub_saturate_u_i16x8: {
14318     unsigned IntNo;
14319     switch (BuiltinID) {
14320     case WebAssembly::BI__builtin_wasm_add_saturate_s_i8x16:
14321     case WebAssembly::BI__builtin_wasm_add_saturate_s_i16x8:
14322       IntNo = Intrinsic::sadd_sat;
14323       break;
14324     case WebAssembly::BI__builtin_wasm_add_saturate_u_i8x16:
14325     case WebAssembly::BI__builtin_wasm_add_saturate_u_i16x8:
14326       IntNo = Intrinsic::uadd_sat;
14327       break;
14328     case WebAssembly::BI__builtin_wasm_sub_saturate_s_i8x16:
14329     case WebAssembly::BI__builtin_wasm_sub_saturate_s_i16x8:
14330       IntNo = Intrinsic::wasm_sub_saturate_signed;
14331       break;
14332     case WebAssembly::BI__builtin_wasm_sub_saturate_u_i8x16:
14333     case WebAssembly::BI__builtin_wasm_sub_saturate_u_i16x8:
14334       IntNo = Intrinsic::wasm_sub_saturate_unsigned;
14335       break;
14336     default:
14337       llvm_unreachable("unexpected builtin ID");
14338     }
14339     Value *LHS = EmitScalarExpr(E->getArg(0));
14340     Value *RHS = EmitScalarExpr(E->getArg(1));
14341     Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
14342     return Builder.CreateCall(Callee, {LHS, RHS});
14343   }
14344   case WebAssembly::BI__builtin_wasm_bitselect: {
14345     Value *V1 = EmitScalarExpr(E->getArg(0));
14346     Value *V2 = EmitScalarExpr(E->getArg(1));
14347     Value *C = EmitScalarExpr(E->getArg(2));
14348     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_bitselect,
14349                                      ConvertType(E->getType()));
14350     return Builder.CreateCall(Callee, {V1, V2, C});
14351   }
14352   case WebAssembly::BI__builtin_wasm_min_s_i8x16:
14353   case WebAssembly::BI__builtin_wasm_min_u_i8x16:
14354   case WebAssembly::BI__builtin_wasm_max_s_i8x16:
14355   case WebAssembly::BI__builtin_wasm_max_u_i8x16:
14356   case WebAssembly::BI__builtin_wasm_min_s_i16x8:
14357   case WebAssembly::BI__builtin_wasm_min_u_i16x8:
14358   case WebAssembly::BI__builtin_wasm_max_s_i16x8:
14359   case WebAssembly::BI__builtin_wasm_max_u_i16x8:
14360   case WebAssembly::BI__builtin_wasm_min_s_i32x4:
14361   case WebAssembly::BI__builtin_wasm_min_u_i32x4:
14362   case WebAssembly::BI__builtin_wasm_max_s_i32x4:
14363   case WebAssembly::BI__builtin_wasm_max_u_i32x4: {
14364     unsigned IntNo;
14365     switch (BuiltinID) {
14366     case WebAssembly::BI__builtin_wasm_min_s_i8x16:
14367     case WebAssembly::BI__builtin_wasm_min_s_i16x8:
14368     case WebAssembly::BI__builtin_wasm_min_s_i32x4:
14369       IntNo = Intrinsic::wasm_min_signed;
14370       break;
14371     case WebAssembly::BI__builtin_wasm_min_u_i8x16:
14372     case WebAssembly::BI__builtin_wasm_min_u_i16x8:
14373     case WebAssembly::BI__builtin_wasm_min_u_i32x4:
14374       IntNo = Intrinsic::wasm_min_unsigned;
14375       break;
14376     case WebAssembly::BI__builtin_wasm_max_s_i8x16:
14377     case WebAssembly::BI__builtin_wasm_max_s_i16x8:
14378     case WebAssembly::BI__builtin_wasm_max_s_i32x4:
14379       IntNo = Intrinsic::wasm_max_signed;
14380       break;
14381     case WebAssembly::BI__builtin_wasm_max_u_i8x16:
14382     case WebAssembly::BI__builtin_wasm_max_u_i16x8:
14383     case WebAssembly::BI__builtin_wasm_max_u_i32x4:
14384       IntNo = Intrinsic::wasm_max_unsigned;
14385       break;
14386     default:
14387       llvm_unreachable("unexpected builtin ID");
14388     }
14389     Value *LHS = EmitScalarExpr(E->getArg(0));
14390     Value *RHS = EmitScalarExpr(E->getArg(1));
14391     Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
14392     return Builder.CreateCall(Callee, {LHS, RHS});
14393   }
14394   case WebAssembly::BI__builtin_wasm_dot_s_i32x4_i16x8: {
14395     Value *LHS = EmitScalarExpr(E->getArg(0));
14396     Value *RHS = EmitScalarExpr(E->getArg(1));
14397     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_dot);
14398     return Builder.CreateCall(Callee, {LHS, RHS});
14399   }
14400   case WebAssembly::BI__builtin_wasm_any_true_i8x16:
14401   case WebAssembly::BI__builtin_wasm_any_true_i16x8:
14402   case WebAssembly::BI__builtin_wasm_any_true_i32x4:
14403   case WebAssembly::BI__builtin_wasm_any_true_i64x2:
14404   case WebAssembly::BI__builtin_wasm_all_true_i8x16:
14405   case WebAssembly::BI__builtin_wasm_all_true_i16x8:
14406   case WebAssembly::BI__builtin_wasm_all_true_i32x4:
14407   case WebAssembly::BI__builtin_wasm_all_true_i64x2: {
14408     unsigned IntNo;
14409     switch (BuiltinID) {
14410     case WebAssembly::BI__builtin_wasm_any_true_i8x16:
14411     case WebAssembly::BI__builtin_wasm_any_true_i16x8:
14412     case WebAssembly::BI__builtin_wasm_any_true_i32x4:
14413     case WebAssembly::BI__builtin_wasm_any_true_i64x2:
14414       IntNo = Intrinsic::wasm_anytrue;
14415       break;
14416     case WebAssembly::BI__builtin_wasm_all_true_i8x16:
14417     case WebAssembly::BI__builtin_wasm_all_true_i16x8:
14418     case WebAssembly::BI__builtin_wasm_all_true_i32x4:
14419     case WebAssembly::BI__builtin_wasm_all_true_i64x2:
14420       IntNo = Intrinsic::wasm_alltrue;
14421       break;
14422     default:
14423       llvm_unreachable("unexpected builtin ID");
14424     }
14425     Value *Vec = EmitScalarExpr(E->getArg(0));
14426     Function *Callee = CGM.getIntrinsic(IntNo, Vec->getType());
14427     return Builder.CreateCall(Callee, {Vec});
14428   }
14429   case WebAssembly::BI__builtin_wasm_abs_f32x4:
14430   case WebAssembly::BI__builtin_wasm_abs_f64x2: {
14431     Value *Vec = EmitScalarExpr(E->getArg(0));
14432     Function *Callee = CGM.getIntrinsic(Intrinsic::fabs, Vec->getType());
14433     return Builder.CreateCall(Callee, {Vec});
14434   }
14435   case WebAssembly::BI__builtin_wasm_sqrt_f32x4:
14436   case WebAssembly::BI__builtin_wasm_sqrt_f64x2: {
14437     Value *Vec = EmitScalarExpr(E->getArg(0));
14438     Function *Callee = CGM.getIntrinsic(Intrinsic::sqrt, Vec->getType());
14439     return Builder.CreateCall(Callee, {Vec});
14440   }
14441   case WebAssembly::BI__builtin_wasm_qfma_f32x4:
14442   case WebAssembly::BI__builtin_wasm_qfms_f32x4:
14443   case WebAssembly::BI__builtin_wasm_qfma_f64x2:
14444   case WebAssembly::BI__builtin_wasm_qfms_f64x2: {
14445     Value *A = EmitScalarExpr(E->getArg(0));
14446     Value *B = EmitScalarExpr(E->getArg(1));
14447     Value *C = EmitScalarExpr(E->getArg(2));
14448     unsigned IntNo;
14449     switch (BuiltinID) {
14450     case WebAssembly::BI__builtin_wasm_qfma_f32x4:
14451     case WebAssembly::BI__builtin_wasm_qfma_f64x2:
14452       IntNo = Intrinsic::wasm_qfma;
14453       break;
14454     case WebAssembly::BI__builtin_wasm_qfms_f32x4:
14455     case WebAssembly::BI__builtin_wasm_qfms_f64x2:
14456       IntNo = Intrinsic::wasm_qfms;
14457       break;
14458     default:
14459       llvm_unreachable("unexpected builtin ID");
14460     }
14461     Function *Callee = CGM.getIntrinsic(IntNo, A->getType());
14462     return Builder.CreateCall(Callee, {A, B, C});
14463   }
14464   case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8:
14465   case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8:
14466   case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4:
14467   case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4: {
14468     Value *Low = EmitScalarExpr(E->getArg(0));
14469     Value *High = EmitScalarExpr(E->getArg(1));
14470     unsigned IntNo;
14471     switch (BuiltinID) {
14472     case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8:
14473     case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4:
14474       IntNo = Intrinsic::wasm_narrow_signed;
14475       break;
14476     case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8:
14477     case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4:
14478       IntNo = Intrinsic::wasm_narrow_unsigned;
14479       break;
14480     default:
14481       llvm_unreachable("unexpected builtin ID");
14482     }
14483     Function *Callee =
14484         CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Low->getType()});
14485     return Builder.CreateCall(Callee, {Low, High});
14486   }
14487   case WebAssembly::BI__builtin_wasm_widen_low_s_i16x8_i8x16:
14488   case WebAssembly::BI__builtin_wasm_widen_high_s_i16x8_i8x16:
14489   case WebAssembly::BI__builtin_wasm_widen_low_u_i16x8_i8x16:
14490   case WebAssembly::BI__builtin_wasm_widen_high_u_i16x8_i8x16:
14491   case WebAssembly::BI__builtin_wasm_widen_low_s_i32x4_i16x8:
14492   case WebAssembly::BI__builtin_wasm_widen_high_s_i32x4_i16x8:
14493   case WebAssembly::BI__builtin_wasm_widen_low_u_i32x4_i16x8:
14494   case WebAssembly::BI__builtin_wasm_widen_high_u_i32x4_i16x8: {
14495     Value *Vec = EmitScalarExpr(E->getArg(0));
14496     unsigned IntNo;
14497     switch (BuiltinID) {
14498     case WebAssembly::BI__builtin_wasm_widen_low_s_i16x8_i8x16:
14499     case WebAssembly::BI__builtin_wasm_widen_low_s_i32x4_i16x8:
14500       IntNo = Intrinsic::wasm_widen_low_signed;
14501       break;
14502     case WebAssembly::BI__builtin_wasm_widen_high_s_i16x8_i8x16:
14503     case WebAssembly::BI__builtin_wasm_widen_high_s_i32x4_i16x8:
14504       IntNo = Intrinsic::wasm_widen_high_signed;
14505       break;
14506     case WebAssembly::BI__builtin_wasm_widen_low_u_i16x8_i8x16:
14507     case WebAssembly::BI__builtin_wasm_widen_low_u_i32x4_i16x8:
14508       IntNo = Intrinsic::wasm_widen_low_unsigned;
14509       break;
14510     case WebAssembly::BI__builtin_wasm_widen_high_u_i16x8_i8x16:
14511     case WebAssembly::BI__builtin_wasm_widen_high_u_i32x4_i16x8:
14512       IntNo = Intrinsic::wasm_widen_high_unsigned;
14513       break;
14514     default:
14515       llvm_unreachable("unexpected builtin ID");
14516     }
14517     Function *Callee =
14518         CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Vec->getType()});
14519     return Builder.CreateCall(Callee, Vec);
14520   }
14521   default:
14522     return nullptr;
14523   }
14524 }
14525 
14526 Value *CodeGenFunction::EmitHexagonBuiltinExpr(unsigned BuiltinID,
14527                                                const CallExpr *E) {
14528   SmallVector<llvm::Value *, 4> Ops;
14529   Intrinsic::ID ID = Intrinsic::not_intrinsic;
14530 
14531   auto MakeCircLd = [&](unsigned IntID, bool HasImm) {
14532     // The base pointer is passed by address, so it needs to be loaded.
14533     Address BP = EmitPointerWithAlignment(E->getArg(0));
14534     BP = Address(Builder.CreateBitCast(BP.getPointer(), Int8PtrPtrTy),
14535                  BP.getAlignment());
14536     llvm::Value *Base = Builder.CreateLoad(BP);
14537     // Operands are Base, Increment, Modifier, Start.
14538     if (HasImm)
14539       Ops = { Base, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)),
14540               EmitScalarExpr(E->getArg(3)) };
14541     else
14542       Ops = { Base, EmitScalarExpr(E->getArg(1)),
14543               EmitScalarExpr(E->getArg(2)) };
14544 
14545     llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops);
14546     llvm::Value *NewBase = Builder.CreateExtractValue(Result, 1);
14547     llvm::Value *LV = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)),
14548                                             NewBase->getType()->getPointerTo());
14549     Address Dest = EmitPointerWithAlignment(E->getArg(0));
14550     // The intrinsic generates two results. The new value for the base pointer
14551     // needs to be stored.
14552     Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment());
14553     return Builder.CreateExtractValue(Result, 0);
14554   };
14555 
14556   auto MakeCircSt = [&](unsigned IntID, bool HasImm) {
14557     // The base pointer is passed by address, so it needs to be loaded.
14558     Address BP = EmitPointerWithAlignment(E->getArg(0));
14559     BP = Address(Builder.CreateBitCast(BP.getPointer(), Int8PtrPtrTy),
14560                  BP.getAlignment());
14561     llvm::Value *Base = Builder.CreateLoad(BP);
14562     // Operands are Base, Increment, Modifier, Value, Start.
14563     if (HasImm)
14564       Ops = { Base, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)),
14565               EmitScalarExpr(E->getArg(3)), EmitScalarExpr(E->getArg(4)) };
14566     else
14567       Ops = { Base, EmitScalarExpr(E->getArg(1)),
14568               EmitScalarExpr(E->getArg(2)), EmitScalarExpr(E->getArg(3)) };
14569 
14570     llvm::Value *NewBase = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops);
14571     llvm::Value *LV = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)),
14572                                             NewBase->getType()->getPointerTo());
14573     Address Dest = EmitPointerWithAlignment(E->getArg(0));
14574     // The intrinsic generates one result, which is the new value for the base
14575     // pointer. It needs to be stored.
14576     return Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment());
14577   };
14578 
14579   // Handle the conversion of bit-reverse load intrinsics to bit code.
14580   // The intrinsic call after this function only reads from memory and the
14581   // write to memory is dealt by the store instruction.
14582   auto MakeBrevLd = [&](unsigned IntID, llvm::Type *DestTy) {
14583     // The intrinsic generates one result, which is the new value for the base
14584     // pointer. It needs to be returned. The result of the load instruction is
14585     // passed to intrinsic by address, so the value needs to be stored.
14586     llvm::Value *BaseAddress =
14587         Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy);
14588 
14589     // Expressions like &(*pt++) will be incremented per evaluation.
14590     // EmitPointerWithAlignment and EmitScalarExpr evaluates the expression
14591     // per call.
14592     Address DestAddr = EmitPointerWithAlignment(E->getArg(1));
14593     DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), Int8PtrTy),
14594                        DestAddr.getAlignment());
14595     llvm::Value *DestAddress = DestAddr.getPointer();
14596 
14597     // Operands are Base, Dest, Modifier.
14598     // The intrinsic format in LLVM IR is defined as
14599     // { ValueType, i8* } (i8*, i32).
14600     Ops = {BaseAddress, EmitScalarExpr(E->getArg(2))};
14601 
14602     llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops);
14603     // The value needs to be stored as the variable is passed by reference.
14604     llvm::Value *DestVal = Builder.CreateExtractValue(Result, 0);
14605 
14606     // The store needs to be truncated to fit the destination type.
14607     // While i32 and i64 are natively supported on Hexagon, i8 and i16 needs
14608     // to be handled with stores of respective destination type.
14609     DestVal = Builder.CreateTrunc(DestVal, DestTy);
14610 
14611     llvm::Value *DestForStore =
14612         Builder.CreateBitCast(DestAddress, DestVal->getType()->getPointerTo());
14613     Builder.CreateAlignedStore(DestVal, DestForStore, DestAddr.getAlignment());
14614     // The updated value of the base pointer is returned.
14615     return Builder.CreateExtractValue(Result, 1);
14616   };
14617 
14618   switch (BuiltinID) {
14619   case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry:
14620   case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry_128B: {
14621     Address Dest = EmitPointerWithAlignment(E->getArg(2));
14622     unsigned Size;
14623     if (BuiltinID == Hexagon::BI__builtin_HEXAGON_V6_vaddcarry) {
14624       Size = 512;
14625       ID = Intrinsic::hexagon_V6_vaddcarry;
14626     } else {
14627       Size = 1024;
14628       ID = Intrinsic::hexagon_V6_vaddcarry_128B;
14629     }
14630     Dest = Builder.CreateBitCast(Dest,
14631         llvm::VectorType::get(Builder.getInt1Ty(), Size)->getPointerTo(0));
14632     LoadInst *QLd = Builder.CreateLoad(Dest);
14633     Ops = { EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), QLd };
14634     llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14635     llvm::Value *Vprd = Builder.CreateExtractValue(Result, 1);
14636     llvm::Value *Base = Builder.CreateBitCast(EmitScalarExpr(E->getArg(2)),
14637                                               Vprd->getType()->getPointerTo(0));
14638     Builder.CreateAlignedStore(Vprd, Base, Dest.getAlignment());
14639     return Builder.CreateExtractValue(Result, 0);
14640   }
14641   case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry:
14642   case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry_128B: {
14643     Address Dest = EmitPointerWithAlignment(E->getArg(2));
14644     unsigned Size;
14645     if (BuiltinID == Hexagon::BI__builtin_HEXAGON_V6_vsubcarry) {
14646       Size = 512;
14647       ID = Intrinsic::hexagon_V6_vsubcarry;
14648     } else {
14649       Size = 1024;
14650       ID = Intrinsic::hexagon_V6_vsubcarry_128B;
14651     }
14652     Dest = Builder.CreateBitCast(Dest,
14653         llvm::VectorType::get(Builder.getInt1Ty(), Size)->getPointerTo(0));
14654     LoadInst *QLd = Builder.CreateLoad(Dest);
14655     Ops = { EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), QLd };
14656     llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14657     llvm::Value *Vprd = Builder.CreateExtractValue(Result, 1);
14658     llvm::Value *Base = Builder.CreateBitCast(EmitScalarExpr(E->getArg(2)),
14659                                               Vprd->getType()->getPointerTo(0));
14660     Builder.CreateAlignedStore(Vprd, Base, Dest.getAlignment());
14661     return Builder.CreateExtractValue(Result, 0);
14662   }
14663   case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pci:
14664     return MakeCircLd(Intrinsic::hexagon_L2_loadrub_pci, /*HasImm*/true);
14665   case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pci:
14666     return MakeCircLd(Intrinsic::hexagon_L2_loadrb_pci,  /*HasImm*/true);
14667   case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pci:
14668     return MakeCircLd(Intrinsic::hexagon_L2_loadruh_pci, /*HasImm*/true);
14669   case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pci:
14670     return MakeCircLd(Intrinsic::hexagon_L2_loadrh_pci,  /*HasImm*/true);
14671   case Hexagon::BI__builtin_HEXAGON_L2_loadri_pci:
14672     return MakeCircLd(Intrinsic::hexagon_L2_loadri_pci,  /*HasImm*/true);
14673   case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pci:
14674     return MakeCircLd(Intrinsic::hexagon_L2_loadrd_pci,  /*HasImm*/true);
14675   case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pcr:
14676     return MakeCircLd(Intrinsic::hexagon_L2_loadrub_pcr, /*HasImm*/false);
14677   case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pcr:
14678     return MakeCircLd(Intrinsic::hexagon_L2_loadrb_pcr,  /*HasImm*/false);
14679   case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pcr:
14680     return MakeCircLd(Intrinsic::hexagon_L2_loadruh_pcr, /*HasImm*/false);
14681   case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pcr:
14682     return MakeCircLd(Intrinsic::hexagon_L2_loadrh_pcr,  /*HasImm*/false);
14683   case Hexagon::BI__builtin_HEXAGON_L2_loadri_pcr:
14684     return MakeCircLd(Intrinsic::hexagon_L2_loadri_pcr,  /*HasImm*/false);
14685   case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pcr:
14686     return MakeCircLd(Intrinsic::hexagon_L2_loadrd_pcr,  /*HasImm*/false);
14687   case Hexagon::BI__builtin_HEXAGON_S2_storerb_pci:
14688     return MakeCircSt(Intrinsic::hexagon_S2_storerb_pci, /*HasImm*/true);
14689   case Hexagon::BI__builtin_HEXAGON_S2_storerh_pci:
14690     return MakeCircSt(Intrinsic::hexagon_S2_storerh_pci, /*HasImm*/true);
14691   case Hexagon::BI__builtin_HEXAGON_S2_storerf_pci:
14692     return MakeCircSt(Intrinsic::hexagon_S2_storerf_pci, /*HasImm*/true);
14693   case Hexagon::BI__builtin_HEXAGON_S2_storeri_pci:
14694     return MakeCircSt(Intrinsic::hexagon_S2_storeri_pci, /*HasImm*/true);
14695   case Hexagon::BI__builtin_HEXAGON_S2_storerd_pci:
14696     return MakeCircSt(Intrinsic::hexagon_S2_storerd_pci, /*HasImm*/true);
14697   case Hexagon::BI__builtin_HEXAGON_S2_storerb_pcr:
14698     return MakeCircSt(Intrinsic::hexagon_S2_storerb_pcr, /*HasImm*/false);
14699   case Hexagon::BI__builtin_HEXAGON_S2_storerh_pcr:
14700     return MakeCircSt(Intrinsic::hexagon_S2_storerh_pcr, /*HasImm*/false);
14701   case Hexagon::BI__builtin_HEXAGON_S2_storerf_pcr:
14702     return MakeCircSt(Intrinsic::hexagon_S2_storerf_pcr, /*HasImm*/false);
14703   case Hexagon::BI__builtin_HEXAGON_S2_storeri_pcr:
14704     return MakeCircSt(Intrinsic::hexagon_S2_storeri_pcr, /*HasImm*/false);
14705   case Hexagon::BI__builtin_HEXAGON_S2_storerd_pcr:
14706     return MakeCircSt(Intrinsic::hexagon_S2_storerd_pcr, /*HasImm*/false);
14707   case Hexagon::BI__builtin_brev_ldub:
14708     return MakeBrevLd(Intrinsic::hexagon_L2_loadrub_pbr, Int8Ty);
14709   case Hexagon::BI__builtin_brev_ldb:
14710     return MakeBrevLd(Intrinsic::hexagon_L2_loadrb_pbr, Int8Ty);
14711   case Hexagon::BI__builtin_brev_lduh:
14712     return MakeBrevLd(Intrinsic::hexagon_L2_loadruh_pbr, Int16Ty);
14713   case Hexagon::BI__builtin_brev_ldh:
14714     return MakeBrevLd(Intrinsic::hexagon_L2_loadrh_pbr, Int16Ty);
14715   case Hexagon::BI__builtin_brev_ldw:
14716     return MakeBrevLd(Intrinsic::hexagon_L2_loadri_pbr, Int32Ty);
14717   case Hexagon::BI__builtin_brev_ldd:
14718     return MakeBrevLd(Intrinsic::hexagon_L2_loadrd_pbr, Int64Ty);
14719   default:
14720     break;
14721   } // switch
14722 
14723   return nullptr;
14724 }
14725